docs: fix factual errors in openhermes-sft example page

[Shi-Dong]

Summary

Fixes factual errors in docs/examples/openhermes-sft.md found during a fact-check of every page under docs/examples/ against the actual code.

• Disk requirement: OpenHermes-2.5 is ~1.6 GB on the Hub (a few GB converted), not "~50 GB". Corrected to ~5 GB.
• What to watch: the page listed fabricated sft/* metrics and a data/prefetch_queue_depth gauge that the code never logs. sft_loss_function (in miles/backends/training_utils/loss_hub/losses.py) reports a single "loss" metric, which appears as train/loss in the per-step log line. Replaced the fake block with the real step N: {'train/loss': ..., 'train/grad_norm': ...} shape.
• LR schedule knob: the doc recommended --lr-warmup-iters 100; the actual launch script scripts/run-qwen3-4B-base-sft.sh uses --lr-decay-style cosine --lr-warmup-fraction 0.1. The table now matches the script.
• Mix datasets variation: the doc showed --prompt-data taking multiple entries and claimed per-source loss is logged separately. --prompt-data is a single-path argument and there is no per-source loss. Replaced with the working approach: concatenate sources into one parquet with datasets.concatenate_datasets before training.

Test plan

• Verified against scripts/run-qwen3-4B-base-sft.sh, miles/backends/training_utils/loss_hub/losses.py, and the --prompt-data definition in miles/arguments.py on current main
• Docs render check

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.