fix(compute): GPUEngine.Reshape honors dst argument (#81)

[dndungu]

Summary

• Fixes GPUEngine.Reshape silently ignores dst argument (compute.Engine contract violation) #81. The zero-copy GPUStorage and Float16Storage fast-paths in GPUEngine.Reshape ignored the caller-provided dst and returned a fresh tensor aliasing the source storage. Callers that passed a pre-allocated dst and discarded the return value silently received stale pre-allocated storage — exactly the failure mode that froze PatchTST GPU training (loss=0.268357 across all epochs) on DGX GB10.
• Fix: when dst is provided, mutate dst[0] to alias the reshaped view (SetStorage + SetShape + SetStrides) and return dst[0]. When no dst is provided, preserve current behavior. New aliasReshapeDst helper centralizes the dst-honoring logic for both Float16Storage and GPUStorage[T] fast-paths.

Evidence (from #79 / zerfoo/zerfoo#371 investigation)

First-batch GPU probe in PatchTST backward, captured on DGX GB10:

post-dFlat-matmul      ptr=0x...628c000 l2=0.0833596 allZero=false
post-dFlat-reshape-dX  ptr=0x...f244000 l2=0         allZero=true
encBwd:entry:dX        ptr=0x...f244000 l2=0         allZero=true

MatMul wrote real gradients into fc.dFlat. The very next call,
engine.Reshape(ctx, fc.dFlat, [1600,64], fc.dX), returned a fresh tensor
aliasing fc.dFlat but left fc.dX (the dst) untouched at its pre-allocated
zeros. Caller discarded the return → encoderBackward saw all-zero input →
every downstream gradient was zero → frozen loss.

Tests

compute/gpu_reshape_dst_test.go:

• TestGPUEngine_Reshape_HonorsDst: pre-allocates dst with a -999.0 poison pattern, runs Reshape with dst, asserts (a) ret == dst, (b) shape becomes [2,8], (c) dst data reflects src (1..16) not poison. This is the regression test for the silent-zero trap.
• TestGPUEngine_Reshape_NoDst: preserves the no-dst fast-path behavior — returns a fresh tensor with correct shape and source data.

Both tests skip when CUDA is unavailable (consistent with other GPU tests in this package).

Test plan

• go build ./... clean
• go vet ./... clean
• gofmt clean
• Full go test ./... ztensor test suite passes on CPU host
• CI green (incl. CUDA-enabled runner if configured)

Refs

• Closes GPUEngine.Reshape silently ignores dst argument (compute.Engine contract violation) #81
• Refs GPU engine dst-output routing: tensors written by kernels return zero via .Data() — freezes PatchTST training on GB10 #79 (closed)
• Refs fix(timeseries): PatchTST GPU convergence — use Reshape return value zerfoo#371 (zerfoo-side workaround)