test: one fallback suite on one in-process vendor

[CMGS]

#86 and #88 each added a crates/handler/tests/ file on model fallback (fallback_admission.rs, fallback_origin.rs), with two different transports, beside four inline fallback tests in handler/src/lib.rs that ran their own TCP vendor (vendor_by_model + fallback_handler).

This folds all eight tests into crates/handler/tests/fallback.rs, on the in-process Vendor stub and config fixture from #86:

• the four inline tests move over with their assertions unchanged; the TCP fixture goes
• the fix: preserve requested model through quota fallback #88 regression runs on the shared models: broken fails upstream, its chain target throttled is over its per-model quota and degrades to the tenant fallback, which either serves (healthy) or has no account (unavailable, so the original chain resumes at healthy)
• Vendor::calls() replaces the repeated vendor.0.load(Ordering::Relaxed)

Production code is untouched. Test count is unchanged (99 in gw-handler); tests are −77 lines and one test binary instead of two.

Validation:

• cargo fmt --all --check: clean
• cargo clippy --workspace --all-targets -- -D warnings: clean
• cargo test --workspace: 617 passed, 0 failed (PG/Redis env-gated suites not run)
• mutation: with swap_to_fallback reverted to the pre-fix: preserve requested model through quota fallback #88 unconditional write, the regression fails in each arm — the served arm echoes throttled instead of broken, and the failing arm ends in a 503 because the chain continues from throttled's empty list