Add upstream-ready BFCL AgentWeave model handler

[sauravsingla]

Adds a separate standard-data BFCL compatibility path for publishing AgentWeave as a routed system/model handler, without reusing the custom v5/v6 distractor augmentation.

What this PR adds:

• integrations/bfcl_upstream/agentweave_router.py: reusable AgentWeave router for BFCL-provided candidate functions, intentionally kept outside the historical frozen agentweave/ scope
• upstream-ready AgentWeaveHammerHandler that subclasses BFCL's existing HammerHandler and routes immediately before prompt construction
• compatibility tests proving the router preserves BFCL function objects and respects the frozen tool budget, without adding NumPy to normal dev CI
• a GitHub Actions compatibility proof against bfcl-eval==2025.12.17
• an upstream submission guide with the required BFCL model_config.py, supported-model, dependency, naming, and evidence-boundary steps

Scientific boundary:

• BFCL questions, evaluator ground truth, and function definitions are not altered
• this path is separate from the BFCL-derived routing-pressure v5/v6 studies
• historical frozen manifests/results are not modified
• no score from this adapter should be called an official BFCL leaderboard score until accepted/reproduced through BFCL's normal publication process

Proposed upstream identity: AgentWeave-Hammer2.1-1.5B / AgentWeave + Hammer2.1-1.5B (Prompt).