feat(cake_msa): add Blackwell minimax sparse attention source kernels

.pre-commit-config.yaml

@@ -14,14 +14,16 @@
 # Note the pre-commit hooks shoule only be used for formatting, but not for linting.
 # For linting consider using CI.
 
-# The kernel_src */src trees are upstream kernel drops that are re-synced
-# periodically (sm100/cutedsl_megamoe, sm90/pull_style_cutedsl_megakernel);
-# excluded from all hooks so formatting/lint churn doesn't block syncs.
+# The kernel_src */src trees and blackwell_msa CUDA snapshots are generated or
+# upstream kernel drops that are re-synced periodically
+# (sm100/cutedsl_megamoe, sm90/pull_style_cutedsl_megakernel);
+# excluded from all hooks so formatting/lint churn doesn't block syncs or alter
+# the source-distributed kernel payload.
 # The shim/ layers next to them are ours and stay checked.
 # The Cake selective-state and Mamba CUDA/host trees, along with frozen KDA
 # fallback sources, are byte-faithful generated exports; their SHA-256 digests
 # are checked by JIT manifests and must not be rewritten.
-exclude: ^(?:csrc/cake_selective_state_update/(?:cuda|host)/|csrc/cake_mamba_ssd_combined/generated/|csrc/kda/training_fallback_pointer_sm_(?:100a|103a)\.cu|flashinfer/moe_ep/kernel_src/(?:cutedsl_megamoe|sm90/pull_style_cutedsl_megakernel)/src/)
+exclude: ^(?:csrc/(?:blackwell_msa/|cake_selective_state_update/(?:cuda|host)/|cake_mamba_ssd_combined/generated/|kda/training_fallback_pointer_sm_(?:100a|103a)\.cu)|flashinfer/moe_ep/kernel_src/(?:cutedsl_megamoe|sm90/pull_style_cutedsl_megakernel)/src/)
 
 repos:
   # Standard hooks