De-weird-ify parts of the implementation of deprecated variables

De-weird-ify parts of the implementation of deprecated variables

Don't transform (SYMBOL-[GLOBAL-]VALUE 'S) to a plain use of S
unless exactly semantics-preserving. It had a bunch of issues:
  - it changes behavior if SYMBOL-VALUE is declared notinline
  - it interacts poorly with transforms elsewhere in the compiler
    that may wish to observe the call to symbol-value
    and do a similar thing (to optimize the call per se)
  - it's too closely tied to knowledge of how the compiler translates
    "plain" uses of defvars by safely going through #'SYMBOL-VALUE.
    i.e. the logic for early/late stage deprecation knows that
    the compiler will ultimately call symbol-value and that a separate
    warning should not be issued in ir1-convert-var

As it happens, the easiest way to avoid warning one too many times
is just use FIND-FREE-VAR to warn since it at most once (per symbol
per component) creates a hash-table entry. This also fixes the problem that
no warning was issued for SETQ and LET bindings of deprecated variables.
This approach allows a subsequent change that was hampered by deprecation
checks in the compiler macros. I also suspect that we should re-question the
decision to represent variables in final deprecation stage as symbol-macros.