[vineet-g]

Summary

Finalizing KV cache quantization failed for models whose layers share a KV cache.

Changes

_move_cache_dequant_to_output now treats a cache op with an empty update as a
read-only fetch:

• Its output is dequantized with the params of the op writing the same slot. A writer and its readers are matched by their non-tensor args (e.g. layer_idx).
• Its empty update is rebuilt directly in the quantized dtype, so it matches the quantized cache. Casting or reshaping a zero-size quantized tensor fails to lower. Dynamic dims are read from the original update via sym_size.