Follow up on IR shrinking tasks

[tybug]

We just migrated the shrinker to the IR in #3962, as a part of #3921. While this was a performance win overall, we took some temporary regressions and left notes for where to improve in the future. This issue tracks those items.

• Improve the Float shrinker with special logic for n >= 2**53, where the float gap becomes > 1 (Migrate most shrinker functions to the ir #3962 (comment))
• Adaptively delete in the Collection shrinker:

  hypothesis/hypothesis-python/src/hypothesis/internal/conjecture/shrinking/collection.py

  Lines 44 to 48 in 54d029d

  	# try deleting each element in turn, starting from the back
  	# TODO_BETTER_SHRINK: adaptively delete here by deleting larger chunks at once
  	# if early deletes succeed. use find_integer. turns O(n) into O(log(n))
  	for i in reversed(range(len(self.current))):
  	self.consider(self.current[:i] + self.current[i + 1 :])

• Fix shrinking of stateful tests with bundles. This is a known regression in the shrinker. Likely has to do with misalignments that were allowed via buffers and not via ir (Migrate most shrinker functions to the ir #3962 (comment))
  • 'Realign' IR trees via the intermediary buffer #4086
• Fix any remaining TODO_BETTER_SHRINK comments in the code
• Update docstring comments around the shrinker, and any related internal docs such as guides/internals.rst

Ambitiously (not necessary to close this issue):

• rewrite lstar.py and DFA shrinking for the IR. Here's the commit that took it out: 8eaa4d4