build(deps): bump tract to ^0.21.15

[czoli1976]

Summary

Bumps libDF's tract pin from ^0.21.4 to ^0.21.15. Picks up the items called out in #682:

• 0.21.6 — first WASM-specific f32 4×4 SIMD kernel (the wasm32 path was scalar before this)
• 0.21.8 — MMM kits for matmul + element-wise binary op optimisations
• 0.21.10 — reduce optimisations affecting modern normalisation layers
• 0.21.11 — slice-bubbling + improved gather with compressed inputs
• 0.21.12 — explicit multithread-mm feature flag in tract-linalg

Why two source patches alongside the manifest bump

tract introduced two non-semver-compatible API changes within the 0.21.x patch line, so this can't be a manifest-only bump:

1. Graph::symbol_table → Graph::symbols (introduced 0.21.6) — 3 call sites in libDF/src/tract.rs.

2. ndarray re-export drift — tract-core re-exports a newer ndarray than the standalone ndarray crate at libDF's current pin. Fixed by routing the ndarray import through tract_core::ndarray (using use tract_core::ndarray::{self, prelude::*, …}; so the self alias keeps every later ndarray::* reference resolving against tract's re-export). Five files: tract.rs, wasm.rs, transforms.rs, wav_utils.rs, bin/enhance_wav.rs. Avoids E0308 type-mismatch errors at the libDF ↔ tract API boundary (Axis, prelude::*, ShapeError).

Both renames are mechanical and bit-equivalent.

Test plan — addresses the validation framing in #682

• Native build: cargo check -p deep_filter --features tract,default-model clean on macOS + cargo 1.95.0. Pre-existing warnings only — no errors, no new warnings.
• WASM build: wasm-pack build libDF --target no-modules --release --features wasm succeeds. Final wasm-opt -Oz output is 9.13 MiB vs 8.28 MiB on the 0.21.4 baseline (+10.3%; expected, MMM kits + first WASM SIMD kernel codegen).
• Numerical quality (WASM): DNSMOS P.835 scores on a 30-clip VoiceBank+DEMAND subset are bit-identical to 0.21.4 — SIG=2.9947 BAK=3.6512 OVR=2.6339 P808=3.5753, all four match to all reported decimals.
• Numerical quality (native CLI): deep-filter binary on the same 30-clip subset produces DNSMOS P.835 scores bit-identical to 0.21.4 — SIG=3.5662 BAK=4.1302 OVR=3.3237 P808=3.8365.
• Native RTF (CLI): 0.184 mean (0.21.15) vs 0.195 mean (0.21.4). The 0.21.x kernel work primarily targets WASM, so native gain is modest and partly cargo-profile noise.
• Cross-version sweep also tested 0.21.5 / 0.21.6 / 0.21.12 — DNSMOS bit-equivalence holds across all of them. 0.21.15 chosen as the latest stable patch in the line.

The validation gates this PR matches #682's ask: DNSMOS-based quality validation on a representative corpus, no audio-quality regression (the concern from the 2023 0.19 → 0.20 revert in #405), conservative caret pin (^0.21.15) so the minimum-version bound stays in 0.21.x.

Caveats

• API breaks within a patch series: the two source patches are required because tract introduced API renames inside 0.21.x without a minor-version bump. The diff is small and mechanical; happy to split into one commit per concern (ndarray refactor / symbol_table rename / Cargo.toml + lockfile bump) if you prefer that shape for review.
• Cargo.lock: included since libDF tracks it. Lockfile churn is dependency-resolution only — no logic changes.
• 0.22.x not attempted in this PR. A future bump to 0.22.x is a larger conversation (additional API surface changes, kernel-kit work in flight at sonos/tract). Keeping this PR scoped to staying within the 0.21.x line.

Refs #682.