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

## 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()`:
```python
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()`:
```python
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.