Alternative fix for #205

[sjakobi]

I asked Codex to come up with an alternative to #282 for fixing #205.

I haven't looked at it very closely myself yet.

Notes from Codex:

• Findings

  - #205 is not caused by the strict Int in SLine at prettyprinter/src/Prettyprinter/
    Internal.hs:1669. The real problem is the eager case x of ... in the Line branch at
    prettyprinter/src/Prettyprinter/Internal.hs:2076, because that forces the tail stream
    while layoutSmart is still exploring alternatives.
  - That also explains why only layoutSmart blows up: its fits predicate keeps walking across
    multiple SLines at prettyprinter/src/Prettyprinter/Internal.hs:1995, while layoutPretty
    stops at the first line break.
  - I checked the old sjakobi/issue205 work:
      - a61681e = “Remove bang on Int in SLine”
      - ee48764 = revert of that change
      - 2a18909 = full revert of #139
  - Bench result with the exact issue reproducer, built -O2 with ghc-9.12.2:
      - current master: n=30 -> 8.55s, n=33 -> timed out at 20s
      - lazy-SLine variant: n=30 -> 8.66s, n=33 -> timed out at 20s
      - full revert of #139: effectively instant
  - So the “make the Int lazy” idea is dead: same asymptotics, same runtime.

  Best Candidate

  I prototyped a targeted fix in /tmp/pp-master-detached/prettyprinter/src/Prettyprinter/
  Internal.hs:1981 without touching your main checkout:

  - build layoutSmart with plain SLine i x again
  - add a linear post-pass dropIndentationOnEmptyLines that only rewrites SLine i x to SLine 0
    x when the transformed tail starts with SEmpty or another SLine
  - keep layoutPretty and layoutUnbounded on the current behavior
  - in layoutSmart.fits, only normalize blank-line indentation in the branch that was already
    going to recurse

  That prototype:

  - fixes #205 immediately: n=30, 33, 40 all measured 0.00s
  - preserves the #139 behavior
  - avoids the overreach of removeTrailingWhitespace postprocessing
  - passes the full testsuite via cabal test prettyprinter:testsuite --with-compiler=ghc-9.12

  Other Options

  - Full revert of #139: simplest, but reintroduces the old trailing-indentation regression.
  - removeTrailingWhitespace after layoutSmart: fast and easy, but too aggressive; it strips
    user-produced trailing spaces too.
  - New SimpleDocStream constructor for deferred indentation: probably the cleanest long-term
    design, but much more invasive.