v0.28.11 — Direct-write trackedRefs (ML OOM fix, 253× improvement)

Fixed

• Core: eliminate intermediate trackedRefs copies in pass encoders — Pass encoders
  write ResourceRefs directly to parent CommandEncoder's slice. No per-pass intermediate
  storage, no copy at End(). Encoder pre-allocates capacity 64.

Benchmark (Born ML pattern)

Scenario	allocs/op	B/op
Baseline (no bind)	14	1.9 KB
100 SetBindGroup	21	5.8 KB
1000 SetBindGroup	25	49 KB

For Born ML (10K dispatches/step, 312 encoders): ~15 MB heap vs 3.8 GB (253× improvement).

Matches Rust wgpu where trackers live in the command encoder throughout recording.