Trim obsolete Metal platform capability glue (#602)

This PR is:

- To trust vLLM 0.27 platform defaults for device capability and
quantization validation.
- To remove Metal’s fake SM80 device capability.
- To delete a test that only asserted the fake capability shape.

vLLM now returns `None` for non-CUDA device capability, and an empty
`supported_quantization` list already means there is no platform-side
quantization filter. The Metal overrides were redundant and the fake
CUDA capability was misleading.

Signed-off-by: Yuan Lik Xun <lxyuan0420@gmail.com>