fix(xla): fail fast on MLXCEL_XLA_QUANT=packed for Metal targets

[inureyes]

Summary

Closes #613 (follow-up to #575). The packed int8 path (MLXCEL_XLA_QUANT=packed, #568: resident ui32 weight + f16 scales, dequantized in the StableHLO graph) compiles for metal-spirv but its prefill invoke faults at runtime.

Root cause

With the improved XLA_CHECK diagnostics (below), the fault is:

metal_device.m:650: INTERNAL; Metal command buffer failed (status 5);
while invoking native function hal.fence.await;
[ 0] bytecode prefill@1:84136

The graph lowers (no iree-compile error) and runs token-exact on the GB10 CUDA runtime, so this is IREE metal HAL / metal-spirv runtime behavior on the large ui32-unpack + dequant + matmul command buffer, not an mlxcel emitter defect. A fused quantized-matmul is upstream IREE work (#574), and the packed/int8 bandwidth win is un-demonstrable on the compute-bound Metal decode regardless (per ADR 0004 and the #575 f16 result). So the correct outcome here is robustness on Metal, not a throughput fix.

Changes

• check_packed_supported(device, packed_active) rejects an active packed build on a metal device at the emit/compile boundary (before iree-compile and the weight upload), with an actionable message pointing at the CUDA / CPU targets. Wired into both compile paths (compile_vmfbs single-seq and the batch XlaRaggedEngine load). Backed by a pure packed_runs_on(device, packed_active) predicate with a unit test.
• XLA_CHECK now prints the full annotated IREE status (iree_status_fprint), so any runtime fault names its cause (op / dispatch / driver message) instead of an opaque status code. This is what surfaced the metal_device.m command-buffer error.
• Docs: ADR 0004 and the mlxcel-xla README updated with the root cause and the guard.

Validation (M1 Ultra)

• MLXCEL_XLA_QUANT=packed + MLXCEL_XLA_DEVICE=metal: Error: MLXCEL_XLA_QUANT=packed is not supported on the Metal target: ... (clean, no command-buffer fault).
• f16 on Metal (no packed): still generates (metal default, token-exact per feat: extend the fused low-precision path to the Metal target #575).
• MLXCEL_XLA_QUANT=packed + MLXCEL_XLA_DEVICE=local-task (CPU): still runs (guard is metal-only, packed unaffected off-Metal).
• Unit tests packed_runs_on_rejects_only_active_packed_on_metal and precision_lowers_on_rejects_only_bf16_on_metal pass; cargo fmt --all -- --check clean.