fix(tui): round Describe's percentiles instead of clipping a digit off

[vyncint]

Today

Found while exercising the 0.1.3 Homebrew build against a 2M-row table. The dashboard's Describe panel showed this:

column   type     min     max     nulls   p25    p50    p99
id       Int64    0       1999999 0       512965 996995 197999

That P99 is below the P50 on the same row. The true value is 1,979,969.24.

approx_percentile_cont renders full precision (1979969.2416513609) and the percentile columns are six characters wide, so ratatui clipped the cell from the right and the cut landed mid-integer. Nothing marked it. P25 and P50 were being truncated too, but their integer parts happen to be six digits, so they looked plausible.

Change

before   id  Int64  0  1999999  0  512965 996995 197999
after    id  Int64  0  1999999  0  500678 989902 1.98e6

• A cell that already fits is left exactly as the query rendered it, so 0.0 and 24.75 are untouched and the existing snapshots do not move.
• Anything longer is rounded to the most decimals that fit, then falls back to an exponent form when even the integer part is too wide.
• Over-long text is marked with an ellipsis (FixedSi… for FixedSizeList(8 x non-null Float32)) instead of being cut silently.
• The widths now live in one DESCRIBE_WIDTHS constant that the cell formatters and the layout both read. Previously they existed only in the layout array, which is why no formatter could know what had to fit — that is the structural cause.

Tests

Three, at three levels:

• wide_percentiles_round_instead_of_losing_a_digit feeds the three real percentile strings and asserts they render in order and within 1% of their values. Against the old clipping it fails with percentiles out of order: 505700 996257 197996.
• no_cell_can_overflow_its_column holds every cell inside every width from 1 to 8 across negatives, subnormals, 1e300 and non-numeric input.
• wide_percentiles_render_in_order_in_the_panel renders the real panel through TestBackend, finds the id row and parses its cells back to check P99 lands above P50.

Also verified by driving the built binary through a PTY against a real 2M-row table, which is where the after-row above comes from. Full workspace tests, cargo fmt --check and clippy -D warnings on the crate are green; the PTY suite and all insta snapshots pass unchanged.

Not done

• The panel is still fixed-width and drops columns on a narrow terminal.
• min/max for a float column can still round when very long.

Both are pre-existing behaviour rather than the truncation this fixes.