[CUDA] Preserve executable export names independently from PTX symbols (#24889)

## Motivation

While testing [Qwen3.8-27B](https://huggingface.co/Qwen/Qwen3.8-27B)
end-to-end with IREE, I found a CUDA executable lookup failure.

## Summary

CUDA executable serialization sanitizes kernel names for PTX, but
previously used that same sanitized name as the IREE executable export
identity.

This breaks modules whose generated dispatch names contain characters
that are not valid PTX symbols, such as `$`. Turbine-exported modules
can contain names such as `forward$async_dispatch_*`: compilation
succeeds, but CUDA executable creation fails at runtime.

This change stores the original IREE export name separately from the
sanitized PTX kernel symbol:

- `kernel_name` remains the sanitized symbol used by
`cuModuleGetFunction`.
- `export_name` preserves the original name used by
`hal.executable.lookup.function`.

`export_name` is an appended optional FlatBuffer field. Updated runtimes
fall back to `kernel_name` when loading older VMFBs that do not contain
it.


## Minimal reproduction

A standalone IREE module with a public function containing `$`
reproduces the failure:

```mlir
func.func @qwen$prefill(%input: tensor<4xf32>) -> tensor<4xf32> {
  %output = tensor.empty() : tensor<4xf32>
  %result = linalg.generic {
      indexing_maps = [
        affine_map<(d0) -> (d0)>,
        affine_map<(d0) -> (d0)>
      ],
      iterator_types = ["parallel"]
    }
    ins(%input : tensor<4xf32>) outs(%output : tensor<4xf32>) {
      ^bb0(%in: f32, %out: f32):
        %scaled = arith.mulf %in, %in : f32
        linalg.yield %scaled : f32
    } -> tensor<4xf32>
  return %result : tensor<4xf32>
}
```

```
iree-compile repro.mlir \
  --iree-hal-target-backends=cuda \
  --iree-cuda-target=sm_120 \
  --iree-cuda-target-features=+ptx87 \
  -o repro.vmfb

iree-run-module \
  --module=repro.vmfb \
  --device=cuda \
  --function='qwen$prefill' \
  --input=4xf32=1
```

## Error
```
NOT_FOUND; function 'qwen$prefill' not found in executable
```
Changing only the public symbol to `@qwen_prefill` succeeds and produces
the expected result.

## Testing

Built `iree-compile` and `iree-run-module`, then ran the repro on CUDA
SM120 with PTX 8.7.
After this change it succeeds:

```text
EXEC @qwen$prefill
result[0]: hal.buffer_view
4xf32=1 1 1 1
```

Signed-off-by: weimin023 <tnwilly@gmail.com>