fix(groq): make _combine_llm_outputs None-safe and non-mutating

[Andrei Boldyrev (abcgco)]

Description:

ChatGroq._combine_llm_outputs merged nested token_usage dicts by storing the first generation's nested dict by reference and +=-ing into it — mutating the caller's llm_output, so callbacks/tracing for generation 0 report inflated nested counts (e.g. cached_tokens). A later None clobbered accumulated numbers, and None followed by a number raised TypeError.

This replaces the merge with a recursive _update_token_usage aligned with the pattern merged for Fireworks in #38751: rebuilds dicts instead of mutating, skips None values, preserves accumulator-only nested keys, seeds first-seen nested dict nodes so they merge instead of colliding with a numeric default, hardens token_usage access with .get(), logs-and-passes-through unexpected leaf types, and types the helper via a recursive TokenUsageTree alias.

Groq deltas from the Fireworks version: numeric leaves are int | float — Groq ships float timings such as queue_time/prompt_time inside token_usage (see the response samples in this module's docstrings) — and type mismatches raise TypeError per this package's ruff TRY004 (Fireworks' config doesn't enforce it).

Issue: Fixes #38659

Dependencies: none

Tests: parametrized None/float handling, an immutability guard (source llm_output unchanged, combined dict not aliased), accumulator-key preservation, first-seen nested-dict seeding, parametrized mismatch raises, unexpected-leaf warn-and-pass-through. pytest tests/unit_tests: 78 passed; ruff check/format, mypy, and lint_imports clean.