fix(evaluator): preserve non-string stream values

[ngoncharenko]

Summary

• preserve a lone non-string SSE value under concat instead of coercing it with str()
• continue concatenating string token deltas in arrival order
• reject mixed string/non-string or multi-frame non-string concat streams as partial results with an actionable error
• regenerate the vendored Python SDK mirror

Root cause

PR #812 changed NAT stream aggregation to concat. Its accumulator converted every matched non-string value to a Python string representation.

For example:

SSE:      data: {"value": {"answer": 7}}
Before:   "{'answer': 7}"
After:    {"answer": 7}

A single structured result is not a token delta, so this patch preserves its original JSON type. Text deltas such as ["15", ".", "0"] still become "15.0".

Validation

• uv run --frozen pytest packages/nemo_evaluator_sdk/tests -q: 1125 passed, 2 skipped
• post-rebase focused inference suite: 75 passed
• staged uv run --frozen pre-commit run: passed
• scoped Ruff, Ruff format, and ty checks: passed
• tools/lint/lint-sdk-vendored.sh: passed
• git diff --check origin/main...HEAD: passed

[github-actions]

Suite	Lines Covered	Line Rate	Branch Rate
Unit Tests	26404/34057	77.5%	61.8%
Integration Tests	15162/32682	46.4%	18.6%