harden(emulator-stream): bounded input + stall watchdog

[dev-josias]

Non-breaking robustness fixes from an adversarial review of the M1/M2/M3 emulator core. Scoped to emulator_stream.rs — no overlap with PR #21 (different file, no signature changes, no Cargo.*).

Fixes

• Bounded input channel (was unbounded_channel): try_send drops latest-wins. Prevents unbounded memory growth from gestures sent during the up-to-30s connect window (review finding Feature/integrate react native support #2).
• Stream stall watchdog: a 10s timeout around the frame recv. If the gRPC stream connects but stops emitting (half-open socket), the producer stops instead of hanging the thread + tokio runtime forever (build(deps): bump the all-dependencies group across 1 directory with 3 updates #3).
• Documented the view_to_device ↔ floem letterbox contract (build(deps): bump tree-sitter from 0.22.6 to 0.26.8 in the tree-sitter group across 1 directory #8).

Deferred (intentionally not here — would break #21's API use or need floem)

• build(deps): bump the all-dependencies group with 4 updates #1 (blocker) producer stop-handle / thread-leak + dual-writer — its real home is the panel Start/Stop lifecycle (feat(windows): build + interactive Android emulator panel (v0.3.0 prep) #21's domain) and it changes the public return type; do it as a coordinated follow-up after feat(windows): build + interactive Android emulator panel (v0.3.0 prep) #21 merges.
• build(deps): bump the all-dependencies group with 27 updates #4 pre-decode-once (changes the signal element type), and floem-side build(deps): bump the wasmtime group across 1 directory with 3 updates #5–build(deps): bump the strum group across 1 directory with 2 updates #7 (texture eviction, per-frame buffer reuse) — need a floem change + re-pin.

The review also caught false positives (the "256-byte row alignment" blocker — write_texture accepts any width; use-after-free claims — wgpu's BindGroup holds the texture Arc), excluded from the fix list.

Sole author: dev-josias.