feat: exact weight-byte accounting from safetensors metadata

[inureyes]

Part of #52

Goal

Compute byte-accurate model weight size before load by reading the safetensors header, with the existing analytical estimate (ModelProfile::total_param_bytes) as a fallback.

Why

The most accurate weight number is what is actually on disk. model.safetensors.index.json carries metadata.total_size (exact sum of all tensor bytes), and a single-file model.safetensors carries a header with each tensor's dtype + shape + byte offsets. mlxcel already opens the index (parse_shard_index in src/lib/mlxcel-core/src/weights.rs) but extracts only shard filenames and discards total_size. The analytical estimate (build_profile_from_json) is only "few % accurate" and currently lives only in the distributed path.

Scope / implementation

• Add a function (e.g. weights::weight_footprint_bytes(model_dir) -> Option<u64>) that:
  • prefers metadata.total_size from model.safetensors.index.json when present;
  • else, for a single model.safetensors, reads the 8-byte header length + JSON header and sums per-tensor byte sizes (dtype itemsize × shape product) without loading tensor data;
  • returns None when neither is available (caller falls back to analytical).
• Reuse / extend the existing index parser in weights.rs rather than adding a second JSON reader.
• Provide a single accessor in the estimation module so callers don't each pick between exact/analytical.

Integration (required for completion — not a standalone helper)

• Wire the exact footprint into the shared estimator used by --recommend-quant (src/execution/quant_advisor.rs) and the new estimator (sub-issue D), so the analytical path is only a fallback.
• At minimum, --recommend-quant's "Model size" line must reflect exact bytes when a safetensors header is available.

Acceptance criteria

• Exact byte count returned for both sharded (index.json / total_size) and single-file (model.safetensors header) layouts; None triggers the analytical fallback path.
• Unit tests with fixture headers / index JSON (sharded + single-file + missing).
• One real-model smoke check: exact footprint is within a few % of ModelProfile::total_param_bytes and ≤ the on-disk weight file size.