Fix subagent tool name normalization(#130)

[Andy963]

Fixes #130

Summary

This PR fixes tool-name handling for subagent.allowed_tools.

Previously, run_subagent() forwarded allowed_tools as raw strings. That meant
model-facing aliases such as schedule_list were not normalized back to runtime
registry names such as schedule.list, which could later surface as a confusing
runtime error like:

invalid_input: tools are required for this operation.

Root Cause

subagent was doing raw set filtering:

set(param.allowed_tools) - {"subagent"}

But Bub exposes tool names to models using underscore aliases, while the runtime
registry keeps dotted names. As a result:

• model-facing aliases were not accepted
• surrounding whitespace was not normalized
• exclude did not go through the same canonical resolution path
• unknown tool names were silently ignored instead of failing early

Changes

• add shared tool-name normalization in src/bub/tools.py
• accept both runtime names and model-facing aliases
• normalize input with strip()
• route exclude through the same canonical normalization path
• fail fast on unknown tool names with an explicit ValueError
• keep the default fallback behavior for allowed_tools=None / []
• update run_subagent() to use the shared resolver before dispatch

Tests

Added/updated regression coverage for:

• runtime names and model-facing aliases
• whitespace normalization
• exclude using the same canonical path
• explicit rejection of unknown tool names
• subagent fallback to all non-subagent tools
• subagent alias resolution before agent.run(...)

Verification

uv run pytest -q tests/test_tools.py tests/test_subagent_tool.py

[frostming]

Please fix the lint error