fix(gemma4): diagnose malformed per-module quant overrides

[inureyes]

Summary

PR #690 added Gemma 4 per-module MLP quantization overrides through a flattened overrides map on QuantizationArgs, but a malformed override value (a non-object value, a wrong-typed/negative/zero group_size, or an unsupported bits width) silently fell back to the root defaults with no indication anything was wrong, which is a debugging hazard for quantization frontier work.

What changed

• QuantizationArgs::quant_params_for in src/models/gemma4.rs now validates the override entry matched by the queried module prefix before applying it, reporting an actionable tracing::warn! naming the offending key for a non-object override value, a wrong-typed or negative group_size/bits, a zero group size, or an unsupported bit width (mlxcel only supports the affine {2,3,4,5,6,8} set), then falls back to the root defaults for that field exactly as before.
• Added two small validation helpers, validate_override_group_size and validate_override_bits, that produce the actionable reason strings used in the warnings.
• Metadata keys that legitimately live alongside per-module overrides in the same flattened map (mode, quant_method, quant_algo, config_groups) are untouched because they never coincide with a real module-prefix lookup, so they are never mistaken for malformed overrides.
• Added a quant_override_diagnostics_tests module in src/models/gemma4.rs covering valid full and partial overrides, absent overrides, a non-object override value, string-typed and negative/zero/unsupported-width numeric fields, sibling metadata keys, and a typo'd prefix that simply never matches.

Test plan

• cargo check -p mlxcel --features cuda --lib
• cargo test -p mlxcel --features cuda --lib models::gemma4::quant (24 tests, all pass, including the 11 new diagnostics tests and the pre-existing quant_scheme_tests)
• cargo test -p mlxcel --features cuda --lib models::gemma4_tests (9 pass, 8 ignored for serial MLX execution as before, including the existing gemma4_root_quantization_preserves_per_module_overrides regression test, confirming valid configs load unchanged)
• cargo clippy -p mlxcel --features cuda --lib --tests (no warnings touching gemma4.rs; the crate has one pre-existing unrelated redundant_closure warning in src/models/sanitize.rs that predates this change)
• cargo fmt -p mlxcel -- --check

Closes #691

[inureyes]

PR Finalization Complete

Summary

• Tests: Verified the 11 tests in quant_override_diagnostics_tests cover every bullet in issue bug(gemma4): diagnose malformed per-module quant overrides #691's acceptance criteria (non-object override value, wrong-typed/negative/zero group_size, unsupported bits width, valid full/partial overrides, absent overrides, sibling metadata keys, typo'd prefix). No gap found; no additional tests added.
• Documentation: Checked docs/*.md (single-source English docs; en//ko/ are build-generated and untouched). Per-module quantization override behavior from PR fix(gemma4): honor per-module MLP quant overrides #690 was never user-documented anywhere, so this diagnostics-only change needs no doc update. CHANGELOG.md is not touched by recent comparable small fix PRs (bug(gemma4/cuda): decode collapses to <pad> after 1-3 tokens (single-query maskless attention + quantized GEMV); mlx_lm decodes coherently #688, fix(gemma4): disable CUDA graphs for decode graph-race collapse (#688) #689, fix(gemma4): honor per-module MLP quant overrides #690, fix(cuda): repack ModelOpt NVFP4 to native MLX layout #692), so it was left as-is.
• Lint/Format: cargo fmt -p mlxcel -- --check is clean. cargo clippy -p mlxcel --features cuda --lib --tests -- -D warnings shows exactly one warning, redundant_closure at src/models/sanitize.rs:290, confirmed identical on main and untouched by this PR's diff (out of scope, left as-is).
• Test suite: Re-ran models::gemma4:: (28 tests) and the new quant_override_diagnostics_tests module (11 tests) locally; all pass.

No source changes were needed; working tree is unchanged from the prior commit. Ready for merge.