[ET-VK] Add per-operator dtype constraints to op_registry

Previously, dtype validation for Vulkan operators was scattered across
individual node-checking functions (e.g., `check_to_copy_node` had inline
float16/float32 checks). This made it difficult to see at a glance which
dtypes each operator supports.

This diff introduces a centralized dtype constraint system:

- `utils.py`: Adds dtype set constants (`FP_T`, `INT_T`, `FP_INT32_T`,
  `FP_INT32_BOOL_T`, etc.) and a `DtypeSetList` wrapper class with
  broadcasting semantics. The `check_node_dtypes()` function validates
  tensor inputs/outputs against allowed dtype sets and returns descriptive
  error messages.

- `op_registry.py`: Extends `OpFeatures` with `inputs_dtypes` and
  `outputs_dtypes` parameters. Each operator registration now explicitly
  declares its supported dtypes. Simplified node-checking functions like
  `check_to_copy_node` since dtype logic is now handled declaratively.

- `vulkan_partitioner.py`: Calls `features.check_dtypes()` early in
  validation, failing fast with a clear skip reason if dtypes are invalid.

This approach improves maintainability by making dtype support explicit and
centralizing the validation logic.

Authored with assistance from Claude.

Differential Revision: [D92740295](https://our.internmc.facebook.com/intern/diff/D92740295/)

ghstack-source-id: 339885885
Pull Request resolved: #17336

Author: ssjia