Fix compute->host visibility in HybridForwardPass MoE path (#2)

[pekkah]

Summary

Partial fix for issue #2 (Hybrid GPU+CPU MoE path producing NaN/garbled output). The
documented repro (-g 1 --tq on Qwen3-Coder 30B-A3B) now decodes coherent text at
~14.7 t/s instead of 0 NaN tokens.

Root cause (Bug 1, fixed)

HybridForwardPass.GpuMoeFfn writes the post-RmsNorm hidden state to a host-coherent
BAR buffer (_gpuPinnedNorm) via RecordComputeCopy, then submits and waits on a
fence so the CPU expert-fallback path can read it via MapPinned. On RTX 4070 Ti,
fence completion alone did not make the compute-shader writes visible to host reads
— MapPinned returned stale data, the CPU fallback consumed bogus normPtr values,
the resulting _cpuFallbackBuf was ~1000x out of range, and the residual carried the
corruption through every later layer as the garbled tokens reported in the issue.

Adding an explicit compute -> host pipeline barrier (SHADER_WRITE -> HOST_READ on
COMPUTE_SHADER -> HOST stages) immediately before EndRecordAndSubmit fixes it. The
new helper is VulkanBackend.RecordComputeToHostBarrier().

Residual issue (Bug 2, still open under #2)

Beyond ~-g 9 GPU layers on Qwen3-Coder 30B-A3B, the prefetcher-cached GPU expert
path still produces wrong output even with the host barrier. Disabling the prefetcher
entirely (forcing all experts through CPU fallback) restores correctness across the
full -g 1..-1 range, which points at the GPU-expert MatMul reading prefetched
weights — most likely descriptor-set reuse across multiple recorded dispatches in
ComputePipeline._reusableDs. That fix is a deeper rework and intentionally out of
scope here.

The CLI guard from PR #5 therefore stays in place by default. Set
SHARPI_ALLOW_BROKEN_MOE_HYBRID=1 to bypass the guard for further work on Bug 2.

Test plan

• Original issue repro: SHARPI_ALLOW_BROKEN_MOE_HYBRID=1 ... -g 1 --tq -p \"Hello\"
  produces coherent decode (was: 0 NaN tokens)
• Sweep -g 1..9 with prefetcher-on: all produce coherent output
• Sweep -g 1..-1 with prefetcher disabled (debug only): all produce coherent output
• Default behavior unchanged: guard still refuses MoE on hybrid path with the
  issue Hybrid GPU+CPU path broken for MoE models (GpuMoeFfn) #2 message
• Full test suite green (excluding 2 pre-existing Vulkan failures unrelated to MoE)

🤖 Generated with Claude Code