fix(python): stop double-counting cached tokens in Usage.tally()

[rootkiller6788]

Summary

Usage.tally() double-counts cached tokens when aggregating usage across messages/turns.

The method currently does:

self.total_tokens += subusage.total_tokens
...
self.total_tokens += subusage.cached_token_count

total_tokens is documented as "Total number of tokens used in the request (prompt + generation)". Cached tokens are a subset of the prompt tokens, so they are already included in total_tokens. Adding subusage.cached_token_count a second time inflates the aggregated total.

Repro

from sec_gemini.models.usage import Usage

total = Usage()
total.tally(Usage(
    prompt_tokens=100,
    generated_tokens=50,
    total_tokens=150,
    cached_token_count=30,
))
assert total.total_tokens == 180  # bug: expected 150

Fix

Remove the redundant self.total_tokens += subusage.cached_token_count line so total_tokens is only accumulated once, and add tests/test_usage.py covering single and multiple usage accumulation (including a cached-token case that would fail before this fix).

Verification

• New unit tests pass locally.
• No change to the separate Usage.cost() pricing logic.