feat(xyz): add attention and sparse attention axes

Fourteen axes drive the attention subsystem from the grid: the diffusers
method, the sdp override chain, the dispatcher kernel, five sdnq attention
knobs and six sparse settings. An axis writes shared.opts.data and rebuilds
the chain itself, since the onchange for these settings runs through the
queue lock a grid cell already holds. SharedSettingsStackHelper saves and
restores every attention setting around the grid, keys an axis introduced
included, and an axis for a setting a backend owns warns when that backend
is not in the active chain.

modules/attention/__init__.py

@@ -1,12 +1,12 @@
 """Attention backends: one scaled_dot_product_attention router over the registered backends, the per-generation context, and the diffusers-side processor and dispatcher setup."""
 from modules.attention.registry import AttentionBackend, AttentionCall, Constraints, Platform, Registry, registry
 from modules.attention.router import Plan, PlanEntry, build_plan, get_plan, install_router, reapply, reapply_options, report
-from modules.attention.dispatcher import set_diffusers_attention, set_attention_dispatcher, hijack_kernels, get_kernel_hijack, get_hf_api_hijack
+from modules.attention.dispatcher import set_diffusers_attention, set_attention_dispatcher, list_dispatcher_backends, hijack_kernels, get_kernel_hijack, get_hf_api_hijack
 from modules.attention import backends, context, debug
 
 __all__ = [
     'AttentionBackend', 'AttentionCall', 'Constraints', 'Platform', 'Registry', 'registry',
     'Plan', 'PlanEntry', 'build_plan', 'get_plan', 'install_router', 'reapply', 'reapply_options', 'report',
-    'set_diffusers_attention', 'set_attention_dispatcher', 'hijack_kernels', 'get_kernel_hijack', 'get_hf_api_hijack',
+    'set_diffusers_attention', 'set_attention_dispatcher', 'list_dispatcher_backends', 'hijack_kernels', 'get_kernel_hijack', 'get_hf_api_hijack',
     'backends', 'context', 'debug',
 ]

modules/attention/dispatcher.py

@@ -94,3 +94,13 @@ def set_attention_dispatcher(pipe):
         log.warning(f'Attention dispatcher: active={prev[0].value} list={backends} target={attn} not found')
     else:
         log.debug(f'Attention dispatcher: active={prev[0].value} list={backends}')
+
+
+def list_dispatcher_backends() -> list:
+    """The kernels diffusers can dispatch attention to, for anything that offers hf_attention as a choice."""
+    try:
+        from diffusers.models import attention_dispatch as a
+        return sorted(b.value for b in a._AttentionBackendRegistry.list_backends()) # pylint: disable=protected-access
+    except Exception as e:
+        log.error(f'Attention dispatcher: {e}')
+        return []