flip vulkan implementation

[futz12]

Add a Vulkan implementation of the Flip layer.

Previously Flip had no GPU path and fell back to CPU whenever a graph
contained a flip. This adds a dedicated Vulkan operator with pack1/pack4
shaders.

Implementation notes:

• Mirrors the CPU flip semantics: dims(1-4) -> axis mapping, negative axis
  normalization, and pack4 lane reversal for channel-axis flips.
• One invocation per element; pipeline local size tuned to the device
  subgroup size.
• Empty-axes input is passed through without launching a shader.

Files:

• src/layer/vulkan/flip_vulkan.cpp / flip_vulkan.h
• src/layer/vulkan/shader/flip.comp, flip_pack4.comp

Motivation: RVC-style audio decoders use torch.flip on the GPU; this
keeps the whole graph on Vulkan instead of stalling on a CPU round-trip.