feat: introduce TxTemplate as an intermediate stage

[evanlinjin]

Description

Closes #57.

The transaction-building API had grown awkward, and a revisit traced the awkwardness back to how anti-fee-sniping (AFS) was bolted on. The concrete symptoms:

• Non-AFS callers paid for AFS. Enabling AFS to live inside create_psbt meant everyone had to reckon with two extra error variants and an extra input, whether or not they used the feature.
• AFS was filed under the wrong concept. AFS isn't a PSBT concern at all — it's just logic that decides a transaction's locktime and input sequences. It had no business living inside PSBT creation.
• PsbtParams was doing two jobs. It carried both bitcoin-transaction-shape fields (version, min_locktime, sequence) and PSBT-specific emission options. Because AFS depended on those tx-shape fields, it was structurally pinned inside the PSBT step and couldn't move out.

The right place for AFS is before PSBT creation — it decides tx fields, then those feed into emission. That mirrors Bitcoin Core, where DiscourageFeeSniping runs as a step before the transaction is built. But AFS couldn't be moved without first untangling the tx-shape fields from PsbtParams — i.e. an architectural change, not a local one.

This PR introduces that change: a TxTemplate stage that sits between Selector and Psbt and owns the fully-resolved tx shape (version, locktime, fallback sequence, per-input sequence, ordering).

Selector::try_finalize() → TxTemplate → (Psbt, Finalizer) | Transaction

What this buys us:

• Concerns are separated. Tx-shape decisions live on TxTemplate; PsbtBuildParams is now emission-only. AFS becomes one ordinary, chainable method on TxTemplate (apply_anti_fee_sniping) that just calls the same public setters anyone else would — it is no longer privileged.
• Non-AFS callers stop paying for AFS. The extra error variants and input are gone from the common path.
• Callers who don't want a PSBT aren't forced into one. TxTemplate can emit a bitcoin::Transaction directly, and PSBT v2 support (#48) now has an obvious home as a future TxTemplate::create_psbt_v2() — same construction logic, different emit step.

A secondary win: with create_psbt no longer needing an RNG, the library drops its rand dependency (only rand_core's RngCore trait remains).

Notes to the reviewers

Split into 5 commits:

1. refactor!: rename Selection to TxTemplate — pure rename, no behaviour change. Sets up commit 2.
2. feat(tx-template)!: route tx-shape decisions through TxTemplate — the meat. Adds the resolved tx-shape fields and their validating setters, splits PsbtParams → PsbtBuildParams, makes create_psbt return (Psbt, Finalizer), and turns AFS into a chainable pre-PSBT step. Closes Introduce TxTemplate as a state between Selection and Psbt #57.
3. refactor!: drop rand dependency from the library — moves rand to dev-deps now that emission no longer needs it.
4. feat: Add single_random_draw selection algorithm — ⚠️ scope creep. Adds selection_algorithm_single_random_draw, which shuffles candidates with a caller-supplied rng and selects until the target is met. Strictly speaking this is unrelated to the TxTemplate refactor; it's here because the downstream wallet PoC (bitcoindevkit/bdk_wallet#502) needs a selection algorithm to drive, and the new "randomness lives at the selection step, candidate construction stays deterministic" split made this a natural home for it.
5. feat: Add InputCandidates::push_must_select / push_can_select — ⚠️ also out of scope. Incremental builders that add an Input to a resolved InputCandidates — to the must-select group, or as an optional can-select group — with outpoint de-duplication (InputCandidates was previously construct-once via new). Like commit 4, it's here only because the wallet PoC needs to push caller-supplied foreign inputs onto a candidate set after resolution.

Both (4) and (5) are independent of the TxTemplate change and added purely for downstream convenience — happy to peel them into their own PR(s) if reviewers would rather keep this one focused.

Two implementation points worth a look:

• apply_anti_fee_sniping goes through the public set_locktime / Input::set_sequence path and .expect()s the results. The expects hold by construction: the locktime path only fires when afs_locktime >= current (both height-based, since AFS early-rejects time-based inputs), so the unit + ≥-CLTV checks always pass; the sequence path only touches taproot inputs without a relative timelock, so the CSV/CLTV-disable check can't fail. Worth a second pair of eyes on whether those invariants really are total.
• TxTemplate::from_parts is pub(crate) — external users construct via Selector / InputCandidates. If hand-rolling a TxTemplate becomes a common ask, exposing it is a one-line change.

Out of scope:

• no_std. Blocked on a bdk_coin_select release: fix: add conditional FloatExt import for no_std builds coin-select#37 is merged but unreleased. Follow-up once 0.4.2 lands.
• PSBT v2. Accommodated by the design (see above) but not implemented here.

Changelog notice

• Renamed Selection to TxTemplate, now the single workspace for transaction shaping (version, locktime, fallback sequence, per-input sequence, ordering, anti-fee-sniping, emission).
• Selector::try_finalize now returns Option<TxTemplate>; InputCandidates::into_selection is now into_tx_template, returning Result<TxTemplate, IntoTxTemplateError> (was IntoSelectionError).
• Split PsbtParams into PsbtBuildParams (emission-only). Tx-shape options moved to TxTemplate setters: set_version, set_locktime, set_fallback_sequence, apply_anti_fee_sniping.
• TxTemplate::create_psbt now returns (Psbt, Finalizer). shuffle_inputs is now consuming (returns Self).
• Behaviour change: previously-silent locktime handling is now explicit. A wrong-unit min_locktime was silently ignored and a below-CLTV value silently clamped; set_locktime now returns SetLockTimeError::UnitMismatch / BelowInputCltv, and set_version returns SetVersionError::RelativeTimelockRequiresV2. The hardcoded ENABLE_RBF_NO_LOCKTIME fallback sequence is now configurable via set_fallback_sequence (default unchanged).
• The library no longer depends on rand (only rand_core, for the RngCore trait).
• Added selection_algorithm_single_random_draw, a single-random-draw coin selection algorithm for use with Selector::select_with_algorithm.
• Added InputCandidates::push_must_select / push_can_select for adding an Input to a resolved InputCandidates (with outpoint de-duplication).

Before submitting

• I followed the contribution guidelines
• This PR breaks the existing API

[evanlinjin]

This is fully AI generated and not ready for review.