feat: add intra line highlights

[daulet]

When left and right lines are similar enough, highlight what changed.

Before:

After:

[umputun]

thx for the PR. heads up - #73 by @rashpile is also implementing intra-line highlights. some key differences between the two:

• #73 (+3366/-169, 45 files) has opt-in toggle (W key + --word-diff flag), dedicated theme color keys (WordAddBg/WordRemoveBg), status bar icon, collapsed mode integration, ~94% test coverage, full docs update. uses sergi/go-diff as a new dependency
• this PR (+374/-10, 6 files) is much smaller footprint, zero new deps (hand-rolled LCS + Levenshtein), cleaner separation in a single file. always-on by design

a few things to address in this approach:

1. reuses SearchBg for emphasis - when search is active on a highlighted line, intra-line markers and search highlights become indistinguishable. needs separate colors
2. naive mode threshold 1.0 - when remove/add counts are equal, lines are always paired regardless of similarity. two completely unrelated lines get noisy intra-line highlights
3. CI failing - exhaustive switch lint error in intraline.go:48, missing ChangeContext/ChangeDivider cases

the small footprint and zero-dependency approach is a real advantage here. the hand-rolled LCS + Levenshtein keeps things self-contained, which I value.

would be great if you and @rashpile could sync up on this - two parallel implementations of the same feature. there might be synergy, e.g. combining the smaller footprint from here with some of the integration work from #73, or joining forces on one PR.

[umputun]

thx for this. merged a combined version in #87 that takes the core algorithm from this PR:

• regex tokenizer ([\pL\pN_]+|\s+|[^\pL\pN_\s]+) splitting into word/whitespace/punctuation tokens
• LCS DP + backtrace for lcsKeptTokens
• byte-offset based matchRange ranges, reusing the existing infrastructure
• zero new dependencies

paired with @rashpile's production integration from #73 (pairing heuristic, 30% similarity gate, dedicated theme colors with HSL auto-derivation, collapsed mode refactor).

few changes on top of your implementation:

• replaced the 0.60 / 1.00 Levenshtein pairing with prefix/suffix scoring (avoids forced pairing on same-count blocks with unrelated content)
• added a 500-byte cap in changedTokenRanges to prevent O(m*n) LCS memory blowup on minified content
• dedicated word-diff bg colors instead of reusing SearchBg (search + word-diff can now coexist without visual collision)

the final version is always-on matching your design. closing this one as its core algorithm shipped in #87.