refactor: unify operator selection behind generic dispatch

Add astrai/extension/dispatch.py: per-family decision tables over
composable Specs, with explicit-strict / implicit-loose resolution,
ASTR_OPS env overrides, profile presets, and explain traces. Migrate
attention (behavior-preserving facade) and rotary onto it; new tests
cover spec algebra, resolution semantics, and spec-vs-supports_call
consistency.

astrai/extension/backend/rotary.py

@@ -1,7 +1,9 @@
-"""Rotary embedding with auto-dispatch to CUDA kernel.
+"""Rotary embedding dispatch (family "rotary").
 
-Single entry point ``apply_rotary_emb(x, freqs_cis)`` — uses the fused
-CUDA kernel when available, falls back to torch complex multiply otherwise.
+Registered rows: the fused CUDA kernel (bf16 CUDA, inference-only) and the
+torch complex-multiply fallback (autograd-safe).  Selection runs through
+the generic dispatcher, so ``op_backend(rotary=...)`` and
+``ASTR_OPS=rotary=torch`` work exactly like for attention.
 
 Layout: x is [batch, seq_len, n_heads, head_dim] (bf16).
 freqs_cis is [batch, seq_len, dim/2, 2] (f32) — [cos, sin] pairs.
@@ -10,16 +12,17 @@ freqs_cis is [batch, seq_len, dim/2, 2] (f32) — [cos, sin] pairs.
 import torch
 from torch import Tensor
 
+from astrai.extension.dispatch import (
+    CallContext,
+    ImplRecord,
+    Spec,
+    register_family,
+    resolve,
+)
 from astrai.extension.loader import is_available
 from astrai.extension.ops.rotary import rotary_emb as _cuda_rotary
 
-_cache = {"available": None}
-
-
-def _cuda_available() -> bool:
-    if _cache["available"] is None:
-        _cache["available"] = is_available("rotary_emb")
-    return _cache["available"]
+_SPEC_CUDA = Spec.cuda_device() & Spec.dtype_in(torch.bfloat16) & Spec.no_grad()
 
 
 def _torch_apply(x: Tensor, freqs_cis: Tensor) -> Tensor:
@@ -33,6 +36,29 @@ def _torch_apply(x: Tensor, freqs_cis: Tensor) -> Tensor:
     return x_out.to(dtype)
 
 
+def _rotary_records() -> list:
+    return [
+        ImplRecord(
+            family="rotary",
+            name="cuda",
+            obj=_cuda_rotary,
+            spec=_SPEC_CUDA,
+            available=lambda: is_available("rotary_emb"),
+            priority=0,
+        ),
+        ImplRecord(
+            family="rotary",
+            name="torch",
+            obj=_torch_apply,
+            spec=Spec.always(),
+            priority=99,
+        ),
+    ]
+
+
+register_family("rotary", _rotary_records, lambda: _rotary_records()[-1])
+
+
 def apply_rotary_emb(x: Tensor, freqs_cis: Tensor) -> Tensor:
     """Apply rotary embedding to x.
 
@@ -43,11 +69,11 @@ def apply_rotary_emb(x: Tensor, freqs_cis: Tensor) -> Tensor:
     Returns:
         [batch, seq_len, n_heads, head_dim] (bf16)
     """
-    if (
-        _cuda_available()
-        and not torch.is_grad_enabled()
-        and x.is_cuda
-        and x.dtype == torch.bfloat16
-    ):
-        return _cuda_rotary(x, freqs_cis)
-    return _torch_apply(x, freqs_cis)
+    ctx = CallContext(
+        family="rotary",
+        dtype=x.dtype,
+        device_cuda=x.is_cuda,
+        grad_enabled=torch.is_grad_enabled(),
+        raw=(x, freqs_cis),
+    )
+    return resolve("rotary", ctx).record.obj(x, freqs_cis)