refactor: rework attention backend resolution

- explicit attn_backend() context wins over ASTR_BACKEND env
- polymorphic available()/supports_call() replace isinstance dispatch
- cache singleton backend instances to avoid hot-path allocation
- training (fwd=None) resolves cuda > flash > torch by capability
- flash dense supports mask-free calls only; masked training falls back to torch

astrai/inference/runtime/executor.py

@@ -205,8 +205,8 @@ class Executor:
         max_q_heads = config.num_attention_heads
         head_dim = config.hidden_size // config.num_attention_heads
         backend = get_backend()
-        self._graph_supported = backend.supports_graph() and CudaBackend.supports(
-            head_dim=head_dim
+        self._graph_supported = backend.supports_graph() and (
+            CudaBackend.available() and head_dim in CudaBackend.HEAD_DIMS
         )
         self._workspace = InferenceWorkspace(
             max_batch_size=kv_cache.max_batch_size,