testing/bench.py: make_param_key does not handle torch.dtype consistently with make_param_id

[pandacooming]

Problem

In tile_kernels/testing/bench.py, make_param_key() and make_param_id() format benchmark parameter dicts differently when a value is a torch.dtype.

make_param_id() — correctly uses _format_value():

def make_param_id(params: dict) -> str:
    parts = []
    for key in params:
        value = params[key]
        parts.append(f'{key}={_format_value(value)}')  # dtype → 'fp16' ✅
    return '-'.join(parts) if parts else 'default'

make_param_key() — does NOT use _format_value():

def make_param_key(params: dict) -> str:
    param_str = ','.join(
        f'{_SHORT_NAME.get(k, k)}={format(v, f">{_WIDTH.get(k)}") if k in _WIDTH else v}'
        for k, v in params.items() if v != None
    )
    # dtype falls through: v is torch.float16 → str(v) = 'torch.float16' (ugly)

Function	Input {'dtype': torch.float16}	Output
make_param_id		dtype=fp16 ✅
make_param_key		dtype=torch.float16 ❌

Expected Fix

Apply _format_value(v) to each value in make_param_key(), consistent with make_param_id(). This also ensures any future types added to _format_value (like dict) are handled uniformly in both functions.