Render rich reprs without pandas, byte-identical to pandas output - #605

The rich repr and _repr_html_ previously built a pandas DataFrame and
let pandas render it. Rendering is now done in-house: a lightweight
numpy-only frame, a ported text formatter (cell formatting with
precision and scientific switchover, column widths, max_rows and
max_columns truncation, terminal fit, multiline wrap), and an html
builder reproducing the notebook table. With pandas absent, users now
get the same rich repr they previously needed pandas for; with pandas
present, output is byte-identical, and pandas display options that
could affect the old output are still honored (max_rows, min_rows,
max_columns, width, expand_frame_repr, precision, max_colwidth,
chop_threshold, colheader_justify, float_format, html.border,
html.use_mathjax).

Known non-reproduced corners, all obscure: CJK east-asian width
handling; pprint_nest_depth and max_seq_items on huge UDT cells;
notebook_repr_html=False, where pandas' _repr_html_ returns None and
the old code embedded the literal string "None" in its place, so the
new table is a fix rather than a regression; and max_colwidth of 3 or
less, where pandas narrows the column labels too and leaves them
ragged while the hand renderer keeps them padded to a uniform width.
That last case needs a deliberate max_colwidth <= 3 (the default is
50) together with column labels of differing digit counts, and it
affects the text repr only, since the html path strips cell padding.

The no-pandas tests now assert the pandas-absent output equals the
pandas-present output instead of pinning a header-only fallback.

Verified against the previous pandas render with a differential that
swaps which formatting module graphblas.core resolves to and compares
repr() and _repr_html_() end to end: 12376 checks over shapes,
density, mask and transpose views, and every dtype including two UDTs;
82320 checks over float and complex cell formatting (precision 0
through 17, the scientific switchover boundaries, chop_threshold,
float_format); 1024 checks over expression and infix-expression reprs
in both autocompute modes; and 49140 checks over pandas display
options crossed with six terminal sizes. The only divergences are the
two described above.

Rendering is also faster, since pandas' formatter machinery is no
longer built per call: repr of a 4x4 Matrix ~850us -> ~90us, of a
10-element Vector ~1.5ms -> ~124us, of a 1e6-element Vector ~9.6ms ->
~1.1ms, and _repr_html_ of a 4x4 Matrix ~373us -> ~53us. These are
best-of-5 with both paths interleaved in one process; the ratios (7x
to 12x) held across repeated runs on a shared machine, the absolute
figures less so.