Offload cleanup (round 5): init/Session boundary refactor

Init becomes a thin shim. Session owns the offload lifecycle.

Net -120 lines across the runtime; the bigger win is that the two
sides now match the boundary they claim to enforce. Before, init
did 500+ lines of "pre-digest the inputs for Session" (walk AOTI
catalog, build dummies, install via AOTI, compute fqn_offsets,
compute pinned_bytes_total, resolve budget, build session_catalog
by field-copying ConstantMetadata into a near-identical
ConstantInfo) and then handed all of that to Session::create,
which re-validated most of it. The two layers were the same layer.

New split:
  * Init: parse payload (parser does all per-FQN validation),
    fail-fast on cuda_graph / shared_stream, cudaSetDevice(0),
    walk_aoti_catalog (helper), coverage check (catalog <-> schedule
    set equality), AOTI <-> payload data_size cross-check (the one
    genuinely cross-source check), fetch _weights_blob, call
    Session::create.
  * Session::create(payload, handle, catalog, weights_blob,
    compute_stream, context): builds pinned_bytes_total, resolves
    the budget from the runtime-spec chain, builds + installs
    dummies via AOTI, computes fqn_offsets, builds the host mirror,
    creates pool + copy stream, allocates pinned constants,
    registers dummies with ProbeRegistry.

Helpers added to the offload runtime:

  * ``walk_aoti_catalog(handle, method_name) -> Result<AOTICatalog>``
    (session.h, defined in session.cpp). Walks the AOTI container's
    constants, hard-failing on folded constants, empty original_fqns,
    duplicate original_fqns, and any missing AOTI symbol. Returns
    the (fqns, internal_names, data_sizes, fqn_to_index) tuple.
  * ``resolve_budget(context, floor, pinned)`` (session.cpp
    anonymous namespace). Public mb spec -> internal byte spec ->
    default = floor + pinned. Returns the resolved bytes plus a
    BudgetSpec describing what the user set (for the UX hint).
  * AOTICatalog struct (session.h).

Removed:

  * ConstantInfo struct (session.h). It was a copy of
    ConstantMetadata with an added data_ptr field; Session looks
    dummies up internally from ``dummies_.dummy_data_ptrs`` now.
  * session_catalog field-copy loop in init (~25 lines): Session
    walks payload.constants_metadata directly.
  * Pre-Session pinned_bytes_total + budget resolution in init:
    Session does both internally.
  * Dummy install + fqn_offsets computation in init: Session does
    both internally.
  * Session::create's pinned_bytes_total parameter (added in round
    4): Session now computes it.

61 of 61 offload tests pass; lint clean.

Authored with Claude.

Author: mergennachin