Fix bytefield accessibility generation: detect attributed macros, filter XLEN bit-position labels

[Maanvi212006]

Contributes to riscv/riscv-isa-manual#1129

Problem

Two bugs in tools/wavedrom_a11y.py surfaced while re-verifying auto-generated alt text on bytefield diagrams:

1. detect_diagram_type used an exact match on [bytefield]. Once a file had already been processed and gained an alt="..." attribute (e.g. [bytefield, ,svg, alt="..."]), the literal [bytefield] no longer appeared in the file, so re-running the tool on an already-processed file caused it to fall through to diagram_type = 'unknown' and lose all bytefield-specific title/desc generation.

2. The auto-generated field-name list included XLEN-relative bit-position labels (e.g. MXLEN-1, HSXLEN, VSXLEN-2) as if they were real field names, alongside genuine fields like Interrupt or Exception Code. These are bit-width markers analogous to plain numeric labels (which were already filtered), just written symbolically instead of as literal digits.

Fix

• detect_diagram_type: match on the [bytefield prefix (consistent with how [wavedrom is already matched) instead of the exact string [bytefield], so the check is robust to attributes being appended.
• build_accessibility_text: extend the field-label filter to also drop labels matching [A-Z]*XLEN[+-]?\d* (covers XLEN, MXLEN-1, SXLEN-1, HSXLEN, VSXLEN-2, etc.), while still preserving legitimate bracketed field names like BASE[MXLEN-1:2].

Testing

Verified via direct calls to extract_content / build_accessibility_text across representative files:

• Re-processed files (sv39pte, pmpcfg, mvendorid) now correctly detect as bytefield and produce "{NAME} register field layout" titles with clean Fields: ... descriptions.
• Swept all 92 bytefield .edn files for stray XLEN-pattern or empty descriptions; confirmed the only remaining XLEN occurrences are legitimate bracketed field names (BASE[MXLEN-1:2], BASE[SXLEN-1:2], BASE[VSXLEN-1:2]) that should be kept.

[Bill Traynor (wmat)]

Much appreciatd.