Instead, import KEEPALIVE(BOX) as a sequence of
KEEPALIVE for object fields (flattened).

Move the importation to its own method.
Add debug output.
Rename a few things.

Basic verification of the new optimization.

It relies on a precise GC.

Running diffs for win-x64, there was exactly one result,
in this test. GC.KeepAlive here is used as a generic
"make sure that there is a use for the variable
beyond an assigment" function. The new optimization
eliminates this use as the struct in question has
no GC fields. Fix the test by using an actual generic
"use me" function.

Will be broken by advanced interprocediral analysis
that is apparently not as far away as one would expect,
but will do for now.

It is possible for a GC to occur just after
the KeepAlive calls and for CheckSuccess to
wrongly conclude the objects weren't kept alive.

If the source is a GT_LCL_FLD, then we would only want
too keep alive the fields it encloses (or, since TYP_STRUCT
local fields are currently not supported, nothing).

The field-by-field approach results in the struct getting
address-exposed anyway, so there is no good reason to bloat
the IR. Just have the address escape directly.