fix: polish CLI table output

[N283T]

Summary

Three visual polish fixes found during live testing of #38:

• extractSnippet normalizes embedded whitespace — search result snippets no longer contain literal \n/\t or double spaces
• Column.max_width caps per-column width; applied to the Kind column (category/item) at 10 so it stops getting squeezed to cat… when the Match column is long
• renderBoxed shrink switched from proportional to greedy (largest-first) — narrow columns are preserved, excess is absorbed by whichever column is widest

Before / After

mmcif-dict search "electron density" (140 cols, fake tty):

Before:

│ cat… │ pdbx_dcc_map                             │ dual map properties such as
	Real Space electron density Correlation Coefficien… │

After:

│ category │ pdbx_dcc_map                                    │ dual map properties such as Real Space electron density Correlation Coef… │

Test plan

• zig build test — 55 pass (52 prior + 3 new: whitespace normalization, max_width cap, greedy shrink)
• uv run --script tests/test_smoke.py — 25 pass (unchanged)
• Manual: search "electron density" and relations atom_site in a fake tty — visible improvement, no regressions

Out of scope

• Description dedent sometimes fails when the source .dic has outlier lines (e.g. last line at col 0); median-indent heuristic is a separate follow-up

[N283T]

Closing unmerged. The tabular output approach doesn't fit mmCIF identifier data (20-40 char IDs don't pack into 80-col terminals without destructive truncation). Rolling back PR #38 entirely in a follow-up revert PR.