[WIP][DeepSeek] integrate torchao _scaled_group_gemm, refactor all group_gemms into base class and class impl for modularity (#1142)

This PR:
1 - integrates the new FP8 rowwise _scaled_group_gemm from torchao. This
is fully integrated and working but is not perf optimized yet.
2 - refactors the group gemm impls via group gemm base class that
requires:
a - init (set activation function)
b - arrange_expert_weights
c - execute (run the group gemm)
d - is_available (confirm requirements available for this group gemm). 

There are now four implementations implemented using this base class:
"torch" - native PyTorch bf16 (cutlass)
"torchao" - triton based FBGemm derivative (bf16)
"torchfp8" - torchao fp8 rowwise
"dsgemm" - DeepSeek fp8 blockwise/groupwise 

all group gemm code has moved to "group_gemms.py". 

Testing:
a - ran all four group gemms using inference.
b - ran 'torch' with training.

Author: lessw2020

torchtitan/experiments/deepseek_v3/generate.py

@@ -127,7 +127,7 @@ def create_model(dist_config: DistConfig):
     model_args.ep_size = dist_config.ep_size
     model_args.num_stages = dist_config.pp_size
     model_args.stage_idx = dist_config.pp_rank
-    model_args.max_seq_len = 16384
+    model_args.max_seq_len = 4096  # 16384
 
     with dist_config.device, dist_config.mesh:
         model = DeepseekForCausalLM(model_args)
@@ -291,7 +291,7 @@ def generate(
 
         # Print progress indicator every 20 tokens
         if rank == 0 and tokens_generated % 20 == 0:
-            print(".", end="", flush=True)
+            print(f"{color.yellow}:{color.reset}", end="", flush=True)
 
     # Print newline after progress indicator
     if rank == 0: