fix(ftw): preserve auxiliary expert banks

[ErikBPF]

Summary

Preserve expert banks that use a different quant format or row geometry from the primary bank set during FTW conversion and load. Without this, converting a heterogeneous expert layout silently writes only the primary banks.

• add an optional auxiliary-bank contract to ExpertBanks
• write auxiliary banks as a nested FTW checkpoint using the existing writer
• record the nested checkpoint and model-layer mapping in parent metadata
• validate that checkpoint and mapping metadata appear together
• recursively load the nested banks and restore their quant format, sources, and layer IDs
• leave existing FTW checkpoints unchanged

Provider-specific Qwen loading is intentionally separate; this PR contains only the generic FTW preservation mechanism.

Validation contract

The PR includes focused writer/loader tests for:

• per-layer nested entries and metadata
• restored quant format and layer mapping
• exact tensor round-trip
• rejection of incomplete auxiliary metadata

pytest -q tests/test_ftw_auxiliary.py

Python compile, Ruff on the focused test, and git diff --check remain the packaging gates.

Evidence boundary

The public runtime PoC branch does not contain an immutable FTW endpoint report or a validator for a GGUF-versus-FTW throughput comparison. Therefore this PR makes no full-model throughput, latency, memory, or output-hash claim. Its claim is limited to conversion/load preservation exercised by the focused tests in this diff.