fix: handle non-string request bodies in error-logging round-trip dump

[mittalpk]

Summary

RoundTrip.generate() (databricks/sdk/logger/round_trip_logger.py) passes request.body directly to _redacted_dump(), which assumes a str. This crashes in two ways, both inside the "unable to parse response" debug-log path itself — so the crash masks the actual API error behind an unrelated Python traceback:

1. Streamed upload (a file-like object passed as data= to requests, e.g. dbutils.files.upload from a local file): requests leaves request.body as the original file object rather than reading it. _redacted_dump's len(body) call then raises TypeError: object of type '...' has no len().
2. Non-UTF8 bytes body: once JSON parsing fails, the fallback path calls body.encode(...) on an already-bytes value, raising AttributeError: 'bytes' object has no attribute 'encode'.

Fixes #1489 (the streamed-upload crash); the non-UTF8-bytes crash is the same root cause (an assumption that request.body is always str) so it's fixed and tested alongside it rather than left for a follow-up.

Fix

Added _request_body_str() to normalize request.body into a string before it reaches _redacted_dump:

• str passes through unchanged.
• bytes/bytearray decode with errors="replace" (matching the existing pattern already used for response.content a few lines below).
• Anything else (a stream/file-like object) is represented with a fixed "[stream body]" placeholder rather than read — reading it here would consume the stream (which may already be at an unknown position after the actual send) with no guarantee its content is UTF-8 text.

Test plan

• Added two cases to tests/test_errors.py::test_get_api_error covering both crash scenarios (streamed body via io.BytesIO, non-UTF8 bytes body), extending the existing fake_raw_response helper with an optional request_body parameter that's passed through to requests.Request(..., data=...).prepare() — this reproduces exactly how requests represents each body type in production, rather than hand-constructing a fake object.
• Both new tests verified to fail against pre-fix code (git stash the fix in round_trip_logger.py, rerun — one raises TypeError, the other AttributeError, matching the two reported/found crash shapes) and pass post-fix.
• pytest tests/test_errors.py: 52 passed.
• pytest tests/ (excluding tests/integration): 2109 passed; the only failures (test_open_ai_mixin.py, 7 tests) are pre-existing in this environment due to the optional openai/langchain packages not being installed, unrelated to this change — confirmed by running that file alone and seeing the same ModuleNotFoundError on unmodified main.
• ruff check / ruff format --check: clean on both changed files.

[github-actions]

If integration tests don't run automatically, an authorized user can run them manually by following the instructions below:

Trigger:
go/deco-tests-run/sdk-py

Inputs:

• PR number: 1497
• Commit SHA: 2aa9554e0abc2a7d08f59a5a178954cd7735ecd9

Checks will be approved automatically on success.