fix: avoid double-counting context token overhead

[Excelius-Wang]

Summary

• Count system and history messages as a single request to avoid duplicating request-level framing tokens.
• Measure tool schema overhead against the actual message list.
• Preserve tools-only counting when no messages exist and cover both paths with regression tests.
• Document the additive token attribution contract and its tools-only exception.

Why

LiteLLM token counts for separately counted message groups are not additive because each call includes request-level framing. The previous implementation added independently counted system and history values and measured tools against a synthetic empty message, so context utilization could be overestimated.

With LiteLLM 1.80.9, the regression fixture for gpt-4o-mini is counted as 72 tokens for the complete request. The previous calculation returned 79. This change returns 72 and keeps the breakdown additive.

The regression test compares count_context() against LiteLLM's complete-request count without hard-coding those values.

Validation

• Token counter tests passed with LiteLLM 1.50.0, 1.80.9, and 1.97.0.
• Full core test suite: 336 passed, 3 skipped.
• Black, Flake8, and git diff --check passed.
• Sphinx HTML build completed with no warnings beyond the existing baseline.