TABLESTYLE field list: measured block structure, token sequence not yet pinned (10 records) #60

Split out of #55, which decoded MLINESTYLE, MLEADERSTYLE,
ACDBDETAILVIEWSTYLE and ACDBSECTIONVIEWSTYLE and declined this one.

ODA v5.4.1 §20.4 has no prescription for TABLESTYLE, so it needs the
joint-boundary derivation. What is already measured (see
ARCHITECTURE.md §7d.4, reproduce with
cargo run --release --example probe_token_scan -- samples/arc_2013.dwg 101):

• Budgets between the end of the common object data and the
  data-stream boundary: R2004 1849, R2010 6836, R2013 6844, R2018 5820.
  R2004 is an order of magnitude smaller than R2010+, so the two are not
  the same record shape — the R2004 files also carry a separate
  CELLSTYLEMAP object.
• Structure on R2010+: a 52-bit header followed by four cell-style
  blocks of 1738 / 1696 / 1696 / 1662 bits. The string stream names
  them: ["Standard", "", "Table", "", "_TITLE", "", "_HEADER", "", "_DATA"].
• Each block ends in six 168-bit border sub-records of the shape
  CMC (44) + 36 bits + BD (66) + 22 bits — six borders being top,
  bottom, left, right and the two inner ones. Before them the block runs
  CMC + 64 bits + CMC + BD + 10 bits + 4 × BD + 2 × BD + 32
  bits.
• Decoded anchors on arc_2013.dwg handle 101: the leading CMC of
  each block is the 0xC8…… "no colour" method, the second is ByBlock
  (0xC1000000), and the doubles are 4.5 / 6 (text heights),
  1.5 (margins) and 1.125 (border double-line spacing).

Why it was declined rather than guessed: each corpus file carries
exactly one TABLESTYLE record, so the joint-boundary constraint has
one instance per band instead of the 24 VISUALSTYLE or 11 view-style
records that made those derivations unique. A 52-bit header alone admits
thousands of parses over B/BS/BL/BD/RC/CMC, and the block boundaries
vary by ±42 / −34 bits, so the block layout has variable-width fields
whose positions are not pinned.

The way in is probably a drawing with more than one table style, or a
cross-version joint search that requires the same token sequence to
close the R2010, R2013 and R2018 records at once (they differ by 8 and
1024 bits respectively). Ten records on the corpus; decoding it would
move the aggregate from 90.3% to roughly 90.7%.