docs: list every value with_high_performance() sets and its memory cost

[7487]

Fixes #926.

Answering the scoping question in the issue thread: this PR takes the xet-core code comment half. The hub docs describe each HF_XET_* variable individually, but the doc comment on with_high_performance() is the one place that defines what the preset itself does, and it summarized twelve assignments in one sentence with no memory figure.

Since #943 the buffer values are memory-derived, so the comment now lists:

• all twelve fields with default -> high performance values, the buffer entries as their derivation formulas (usable/16 -> usable/8 etc.) with floors/ceilings, plus the static fallbacks used when memory cannot be probed;
• the worst-case buffer allocation, download_buffer_size + n_active * download_buffer_perfile_size capped at download_buffer_limit (verified against FileReconstructor's target formula) — up to 32GB at the ceilings with the default 8 concurrent downloads;
• that data.max_concurrent_file_downloads is not raised by the preset, only its ingestion twin;
• that every field has its own HF_XET_* variable, so parts of the preset can be adopted individually on memory-constrained machines.

All values checked against the field defaults in config/groups/ and the constants in utils/system_memory.rs on current main.

Comment-only diff; no Rust toolchain on this machine so I did not run cargo locally, but no code is touched and the [floor, ceiling] notation matches the existing field docs in reconstruction.rs.

Written with Claude Code, values verified against the source as described above.

---

Note

Low Risk
Comment-only change with no code or configuration logic modified.

Overview
Expands the with_high_performance() doc comment so the high-performance preset is documented in one place instead of a single vague sentence.

The comment now lists twelve fields with default → preset values (ingestion concurrency, adaptive concurrency bounds, reconstruction fetch sizes, and memory-derived download buffer formulas with clamps). It explains how usable memory is computed, static 16GB / 2GB / 64GB fallbacks when probing is disabled or fails, and the worst-case download buffer formula (up to ~32GB at ceilings with default concurrent downloads). It also notes that max_concurrent_file_downloads is not raised by the preset, and that each field can be tuned via its own HF_XET_* env var without enabling the full preset.

No runtime behavior changes — documentation only on xet_config.rs.

^{Reviewed by Cursor Bugbot for commit 49fb4e8. Bugbot is set up for automated code reviews on this repo. Configure here.}

[seanses]

There's no need to repeat the code in the comment.