Currently, the lambda rewriter has an early optimization pass in
analysis that tries to find all local functions that only capture 'this'
and remove references to local functions that do the same. There are two
problems with this approach:

    1) Generally, removing information from the tree is a bad idea
    because it hurts further analysis passes that may have needed that
    information.

    2) The optimization strategy itself is very tricky and has a number
    of complex corner cases. This has lead to bugs, for example #19033.

This PR deletes the current method and adds a new optimization routine
at the end of the analysis, operating on assigned scopes and
environments rather than removing captured variable analysis. The new
optimization is as follows: if we end up with an environment containing
only 'this', the environment can be removed, all containing methods can
be moved to the top-level type, and all environments which capture the
'this' environment can instead directly capture the 'this' parameter.
This produces almost the same results as the previous optimization, but
is easier to validate as an algebraic equivalence.

The baseline changes come from the new optimization being less aggressive 
about moving functions which only capture 'this' to the top level. This
appears to be a wash -- some codegen gets slightly better, some gets
slightly worse.

Fixes #19033
Fixes #20577