[varlen Kernel] Add test spec for unified varlen paged attention (#147)

## Goal

This is the spec to smoothly migrate v1 kernel (only paged) to v2 kernel
(paged + varlen), without break anything on main, but enable us to
support continuous batching.




## Summary

Adds `test_metal_unified_attention.py`, adapted from vLLM upstream's
`test_triton_unified_attention.py`.

Three test functions form a triangle validation:

```
        v1 kernel (production)
           /        \
  PASS    /          \   xfail
 (runs   /            \  (v2 not built)
  now)  /              \
       v                v
  ref_paged_attn ----> v2 kernel (unified varlen)
     (naive MLX)  xfail
```

- `test_v1_kernel_vs_reference` - v1 == ref (12 cases, runs now,
validates the reference)
- `test_metal_unified_attn_decode_only` - v2 == v1 (24 cases, xfail,
proves v2 is a drop-in replacement for decode)
- `test_metal_unified_attn` - v2 == ref (192 cases, xfail, full varlen
with mixed prefill+decode, GQA, sliding window, soft cap)

## Migration plan

```
Step 1: Build v2 kernel, test decode-only (q_len=1)
        Scaffolding test goes green: v2 == v1
                    |
Step 2: Extend v2 to handle varlen (q_len > 1)
        Full test goes green: v2 == ref
                    |
Step 3: Delete scaffolding test, replace v1 with v2 in production
```

---------

Signed-off-by: ran <hzz5361@psu.edu>

Author: WindChimeRan