Support symbolic tensor shapes and dynamic dimensions

🚀 Feature

Thunder needs to support tensors with symbolic dimensions in their shapes, where one or more dimensions are not known at compile time but are tracked symbolically.

Current Behavior

Thunder likely concretizes shapes at compile/trace time.

Expected Behavior

Support tensor shapes with symbolic dimensions:

• "bf16[1, s0, 5120]" - batch size concrete, seq_len symbolic
• "i64[s0]" - fully symbolic shape
• Operations preserve symbolic dimensions through computation

Example from torch.compile

to: "i64[1, s50]" = l_args_0_.to(device(type='cuda', index=0))
inputs_embeds: "bf16[1, s50, 5120]" = torch.nn.functional.embedding(to, ...)
batched_outputs_2: "b8[s50, 1056]" = torch._functorch.predispatch._remove_batch_dim(...)

Minimal Reproduction Case

import torch
import thunder

@thunder.jit
def process_variable_length(input_ids: torch.Tensor) -> torch.Tensor:
    """
    Process input with variable sequence length.
    input_ids: [1, seq_len] where seq_len is symbolic (s50)
    """
    batch_size, seq_len = input_ids.shape  # seq_len should be symbolic
    
    # Embedding lookup - output shape should be [1, s50, 5120]
    embedding_weight = torch.randn(50000, 5120, device='cuda', dtype=torch.bfloat16)
    embeds = torch.nn.functional.embedding(input_ids, embedding_weight)
    
    # Operations should preserve symbolic dimension
    # embeds.shape = [1, s50, 5120]
    normed = embeds / embeds.norm(dim=-1, keepdim=True)
    
    return normed

for seq_len in [32, 128, 256]:
    input_ids = torch.randint(0, 50000, (1, seq_len), device='cuda')
    output = process_variable_length(input_ids)
    assert output.shape == (1, seq_len, 5120)

print(process_variable_length._lc_cs.last_epilogue_traces[-1])
# def epilogue(normed):
#   # normed: "cuda:0 bf16[1, 256, 5120]" <---- concrete shape but should be symbolic!
#   return normed

Shape Propagation Rules

1. Input tensors can have symbolic dimensions
2. Operations propagate symbolic dimensions:
   • [s0, 5120] @ [5120, 128] → [s0, 128]
   • [1, s0, 5120].mean(-1) → [1, s0]
3. Reshaping with symbolic dims: .view(-1, s0, 128)

Success Criteria

• Thunder accepts tensors with symbolic shape dimensions
• Symbolic dimensions are tracked through operations
• Shape inference works with symbolic dimensions
• Generated code adapts to runtime shapes
• Integration with symbolic scalar values (https://github.com/Lightning-AI/lightning-thunder/issues/2735)

Start with the minimal reproduction in the issue and inspect the generated epilogue for inputs with sequence lengths 32, 128, and 256. Trace how symbolic dimensions would propagate through embedding, normalization, matrix multiplication, reduction, and reshape operations; done means runtime-generated code preserves symbolic shapes and integrates with issue #2735.