fix(drafter): validate quantization against SUPPORTED_AFFINE_BITS #1790

Problem

Found during review of PR #1778, which added SUPPORTED_AFFINE_BITS, SUPPORTED_AFFINE_GROUP_SIZES and two validators to mlxcel_core::layers and applied them to split-mtp. Two drafter paths still bypass them.

1. The Qwen 3.5 MTP drafter quantizes its dense projections at load with bits and group_size taken straight from its config. quantize_weights_with_mode is a non-Result bridge, so an unsupported value throws inside MLX and aborts the process, the failure PR fix(split-mtp): fix refusal text, --force shards, --q-bits validation #1778 fixed for split-mtp. group_size: 0 panics earlier, on the % at model.rs:265. The drafter loads lazily in the batch worker at the first speculative request, so either takes the server down.
2. The Markov (DFlash) head checks a quantized factor's packed width against its own list [2, 3, 4, 6, 8], which omits 5, so a 5-bit factor is refused at load ("no supported bit depth") although MLX supports 5.

Evidence

• src/lib/mlxcel-core/src/drafter/qwen3_5_mtp/model.rs:243 quantize_dense_projections: config values at :248, % group_size at :265, ffi::quantize_weights_with_mode(w, group_size, bits, "affine") at :270; called from from_path at :212.
• The bridge returns UniquePtr<MlxQuantizedWeights>, not Result (src/lib/mlxcel-core/src/lib.rs:2754-2759).
• validate_bounds (src/lib/mlxcel-core/src/drafter/qwen3_5_mtp/config.rs:169-236), the drafter's config gate, does not check quantization.bits or group_size (:85-95).
• src/lib/mlxcel-core/src/drafter/dflash/markov.rs:200 (const SUPPORTED_BITS: [i64; 5] = [2, 3, 4, 6, 8];) in validate_markov_factor (:162).
• src/lib/mlxcel-core/src/layers.rs: validate_affine_quantization_bits (:1370), validate_affine_quantization_group_size (:1401), SUPPORTED_AFFINE_BITS (:1455), SUPPORTED_AFFINE_GROUP_SIZES (:1464).

Proposed fix

• Call both shared validators in quantize_dense_projections once candidates is known to be non-empty (so before the %), return Result, and map the error to DrafterError::Config in from_path. Validate there rather than in validate_bounds: a pre-converted checkpoint whose tensors this function skips should not be refused over values it never uses.
• Derive Markov's list from SUPPORTED_AFFINE_BITS.
• Update the Used by: lists in layers.rs (:1369, :1400, :1452-1454). The two validators' lists already omit their CLI caller, preflight_checks in src/commands/split_mtp.rs:181-184 (added by PR fix(split-mtp): fix refusal text, --force shards, --q-bits validation #1778), so add it together with the drafter callers.

Acceptance criteria

• An unsupported drafter bits or group_size (including 0) makes the load return an error instead of aborting or panicking (test).
• A 5-bit quantized Markov factor passes VanillaMarkovHead::from_weights, and an unexplainable width is still refused (tests).

Verification

cargo test --workspace --profile test-fast --features metal,accelerate
cargo clippy --workspace --all-targets -- -D warnings
cargo fmt --all -- --check

Each new test must fail with its fix reverted.