bch: forward txs on the send path via bch::new_tx_hashes accessor (#905)

[frstrtr]

What

send_shares() (src/impl/bch/node.cpp) decided which peer-referenced txs to forward by probing requires { obj->m_new_transaction_hashes; } directly on the share object. That flat spelling is false for every BCH share variant:

variant	version	new-tx list
bch::Share	v17	m_tx_info.m_new_transaction_hashes (nested)
bch::NewShare	v33	m_tx_info.m_new_transaction_hashes (nested)
bch::SegwitMiningShare	v34	none (SegWit-less, no m_tx_info)
bch::PaddingBugfixShare	v35	none
bch::MergedMiningShare	v36	none

So the entire needed_txs / remember_tx forwarding block was unreachable — BCH forwarded no tx bytes, and a relayed v17/v33 share referencing a tx the peer lacked was exactly the canonical "referenced unknown transaction" disconnect (#905). The receive side (protocol_actual.cpp, protocol_legacy.cpp, node.hpp) already spells m_tx_info.m_new_transaction_hashes correctly, which is why only the send side was dead.

Fix

New src/impl/bch/share_tx_refs.hpp — bch::new_tx_hashes(obj) returns the nested list for v17/v33 and nullptr for v34+ (nothing to forward). Direct port of ltc::new_tx_hashes (#873, ced0d0b); BCH variant topology is identical. send_shares() now routes through it. A new variant is either handled in one of the two known spellings or reported as "no list" — never silently skipped on a member-name mismatch again.

seam-2 (post-broadcast lock discipline) is a STRUCTURAL NO-OP for BCH

LTC #873 carried a second seam: send_shares() was made to return the hashes it actually wrote, and broadcast_share was reworked so the tracker lock is released before the mining-notify / broadcast fan-out (its ingest path remember_tx inserts into m_known_txs under a different discipline, so reading that map mid-broadcast under the held lock was the hazard #873 closed).

BCH does not need that change, and it is not ported here — established by inspection of the BCH broadcast path, not by analogy:

• BCH's submit/broadcast entrypoint already calls lk.unlock() on the tracker lock before the broadcast + mining_submit fan-out (pool_entrypoint.hpp:561); the broadcast runs on the io thread with no share-tracker lock held. The held-lock-across-broadcast ordering that ltc/doge: gate share broadcast on tx completeness, mark only what was sent #873 had to fix never existed on BCH.
• BCH send_shares() is void and needs no return value: the LTC return-hashes plumbing existed only to let the reworked broadcast_share mark what was actually sent under the new lock ordering. With no ordering to change, there is nothing to report back.

i.e. the invariant #873's seam-2 established for LTC is one BCH already holds structurally. This PR is the seam-1 port only; there is no BCH seam-2 to land. (DGB shares the same seam-1 gap on its own lane and must check its own seam-2 separately; LTC is done via #873.)

Compatibility

Share wire-format unchanged — this is relay plumbing over the existing nested list. p2pool-merged-v36 surface: none.

Test

bch_share_tx_refs_kat_test (new) pins:

• compile-time topology — flat spelling absent on every variant (== the old probe was dead), nested present exactly on v17/v33, accessor null-branch for v34+;
• runtime delivery — v17/v33 list forwarded by reference, v34+ forward nothing.

Registered in src/impl/bch/test/CMakeLists.txt (BCH_ABLA_TESTS) and both COIN_BCH build.yml --target lists (main + AsAN legs). Local: bch_ suite builds clean, c2pool-bch links, KAT (ctest #326) PASS.

Merge held for integrator (push-approval card).