perf(bb/msm): stream-walker accumulate — cache operands/dx across the three inner passes

[AztecBot]

What

Pure implementation-level optimization to the stream-walker accumulate kernel (barretenberg/ts/src/msm_webgpu/wgsl/cuzk/ba_stream_walker.template.wgsl).

Each inner-loop iteration walks the S slots three times (forward prefix → batched inversion → backward peel). The base code re-derived every slot's operands in all three passes: it loaded point_x/point_y/l0_index and recomputed dx = fr_sub(p_rx, p_lx) in the forward pass, reloaded+re-subtracted in the inverse pass, and reloaded a third time in the peel. The cursor isn't advanced and acc_x/acc_y aren't written until the peel, so the operands are identical across all three passes. This commit loads them once in the forward pass into per-slot private caches (op_lx/op_ly/op_rx/op_ry, dx_cache) and reuses them — removing ~1/3 of the kernel's global point reads and one field subtraction per slot per iteration.

• Caches live in private registers (same class as the existing acc_x/acc_y) → no GPU memory change.
• Algorithm unchanged / intended bit-identical: no GLV, no signed-digit change, batched-inversion math untouched. S=8, WALKER_TPB=64, pref_scratch (16 KB workgroup), window/bucket sizes and all buffers unchanged.
• Confirmed a real change vs base: dx_cache is absent in stream-walker-impl@eaa6d3d and present after this commit.

Validation status — NOT VALIDATED (blocked). Do not merge on perf claims.

Real-hardware validation could not be completed in this session. There are currently no trustworthy timing or correctness numbers for this PR:

• Device correctness (noble cross-check): not obtained.
• Apple / Adreno / Mali timing vs baseline: not obtained.

Blockers:

• BrowserStack mobile (Android) workers created via the available MCP tooling are served as a non-WebGPU "Android Browser" (no navigator.gpu), so Mali/Adreno could not run the WebGPU harness at all.
• The screenshot-based result readout (/5/worker/<id>/screenshot.json) did not yield a usable screenshot for the harness; the shared 2-seat pool was saturated and workers stayed queued.

Any Apple/Adreno/Mali timings or "cross-check GREEN" statements in earlier revisions of this description are retracted as unverified.

Candidate follow-up (NOT in this commit)

A second, independent pure-engineering lever was prototyped locally but is likewise unvalidated: move pref_scratch from var<workgroup> (16 KB) to per-invocation private storage — it is single-thread-owned (no cross-thread sharing, no barrier), so this frees the 16 KB workgroup allocation the swarm identified as the #1 occupancy bottleneck, at the cost of S*2 vec4 of private space per invocation. Bit-identical (memory placement only). Needs the same real-hardware validation.

TODO before "done"

• Noble cross-check GREEN on real hardware (small logN).
• Baseline-vs-PR wall-clock timing on ≥1 Apple, ≥1 Adreno, ≥1 Mali device; no memory regression.
• Obtain WebGPU-capable mobile BrowserStack access + a working result readout; paste real numbers.

Build: WGSL is inlined at build time; _generated/shaders.ts is regenerated and included in the diff.

Note: git add -A during PR creation also swept in two dev-harness scratch files (dev/msm-webgpu/bench-entry.ts, bench-banner.ts); they are measurement scaffolding, not part of the kernel change.