feat(core): Dependency Graph

[potatomashed]

Here’s a draft GitHub PR description you can drop into your merge request:

This PR introduces a torch.fx.Graph style doubly‐linked dependency graph

• Nodes wrap an opaque "statement" plus its input/output variables
• Maintains:
  • a stmt_to_node map to find a node by its statement
  • a var_to_producer map from each variable to the node that produces it
  • a var_to_consumers map from each variable to all nodes that consume it
• Provides safe, consistent APIs to:
  • create, insert (before/after), replace, erase nodes
  • query producers/consumers by node or by variable

Why it matters

• Enables IR-level transformations and analyses that need to track data dependencies
• Guarantees correctness of producer/consumer relationships as the graph is edited
• Lays the groundwork for scheduling passes (topological sort, liveness, CSE, fusion, etc.)

Unittests

50 pytest cases covering:

• basic chain graphs, branching graphs, empty/inputs‐only graphs
• all API calls, success & failure modes (invalid anchors, duplicate stmts/vars, head/leaf errors)
• identity vs. equality, iteration‐stop, idempotent clears, and more