fix(webgpu): release GPU resources on failure

[CharlesHoskinson]

What changed

• guarantee sparse Pippenger buffers are returned to the pool or destroyed on both success and failure
• clean window-reduction and readback intermediates when a dispatch or mapAsync fails
• make context shutdown destroy the GPUDevice and prevent reuse of a closed buffer pool
• add failure-injection tests for dispatch, mapping, successful cleanup, and idempotent context shutdown

Why

Browser proving allocates large GPU buffers. Previously, cleanup happened only after a successful readback, so WebGPU validation, shader, or device-loss errors could retain those allocations and make fallback or retry unsafe.

Validation

• npm run lint
• npm test (5 tests)
• GOOS=js GOARCH=wasm go build -buildvcs=false ./backend/accelerated/webgpu/...
• go test -short ./... was also attempted; the WebGPU/backend packages passed, while the full run hit two checkout-environment issues: WSL could not resolve the Windows worktree gitdir for TestVersion, and Solidity golden files differed only by CRLF/LF line endings.

Light Reading

Have a look at my repo for more context. I have a full implementation running in WASM https://github.com/CharlesHoskinson/proof-zk-recovery