fix: assorted correctness bugs across utils, data, models, and tooling

[xy200303]

This PR bundles several small, independently verifiable bug fixes (one commit per area, easy to cherry-pick):

• utils/logging: _get_logger_level caught ValueError, but getattr(logging, name) raises AttributeError for an unknown level name, and the handler never assigned the fallback — an invalid ANGELSPEC_LOG_LEVEL crashed at import time. Now catches AttributeError and falls back to WARNING.
• utils/usp: split_usp_batch assigned a local variable named usp_chunk_size, shadowing the module-level function and making every call raise UnboundLocalError. Renamed the local to chunk_len.
• utils/profiling: with the default profile_step_start=0, profile_step_end=0, the torch profiler schedule got active=0 and tripped torch's schedule assertion. The step range is now validated up front with a clear error message.
• data/parse: the registered deepseek-r1-distill template has end_of_turn_token=None, so GeneralParser.parse crashed with TypeError on re.escape(None) during dataset tokenization (the existing |$ alternative already covers a missing end token). Also _ensure_pad_token no longer clobbers a legitimate pad_token_id of 0 (truthiness check → is None).
• data/utils: the Hub dataset path dropped every column outside a small keep-list, silently discarding per-sample tools / reasoning_effort fields that dataset.py threads into the chat template. Also .csv / .tsv / .txt are advertised as supported local extensions but were fed to the JSON loader; they now route to the csv / text loaders.
• data/dataset: the tokenized-dataset cache key omitted prompt_key, so re-running the same data file with a different prompt_key silently reused a stale cache.
• models/draft: the linear and dynamic RoPE-scaling branches in Llama3-Eagle3 _init_rope, and the linear, dynamic, and yarn branches in DeepSeek-Eagle3 _init_rope, omitted base=, silently defaulting to 10000 while every sibling branch passes the configured rope_theta. A draft config combining these scaling types with rope_theta != 10000 (e.g. Llama-3's 500000) would train with the wrong rotation frequencies.
• transfer/mooncake: the last_hidden_states tensor spec on GET looked up dtypes["hidden_states"] instead of dtypes["last_hidden_states"]; if the two are ever stored with different dtypes the receiver computes the wrong byte size.
• examples/tools: both Hy3 example scripts passed training.num_nodes, which is not a schema field (the field is training_num_nodes), so the CLI override crashed at startup with ConfigKeyError. tools/generate_data.py read resp after the message loop even though it is only assigned for user turns — a conversation with no user message crashed with UnboundLocalError; it now returns an error record like the other validation failures.

Verification

• ruff check clean (also fixed the pre-existing F823 in utils/usp.py it flagged).
• Test suite: 150 passed / 1 skipped; the 24 failures are identical to main and purely environmental (missing mooncake package, and torch.compile needs MSVC on this Windows box).
• Reproduced the split_usp_batch UnboundLocalError and the invalid-log-level crash before/after the fix.