Fixes the bug tickled by the tests in this commit's parent.

These tests tickle an error serializing emtpy messages as fields of
other messages when using speed-optimized serializers.

This ensures that all pathnames for all .proto files we load get put
in PROTO-IMPL::*ALL-SCHEMAS*, not just those we import.

Don't override file extension for imports; assume the string that was
used to form the pathname is accurate. This makes base-path and
proto-file the same, so fold them both into proto-file.

Drop bogus checks for type "proto".

* FIND-SCHEMA no longer works on a string or keyword. Both of these
  were based on the PATHNAME-NAME of the protobuf, which has a high
  probability of collision. (e.g., common.proto -> "COMMON", :COMMON)

  Instead, look up schemas based on package and name. For instance,
  the protobuf foo.proto, declaring "package FooPackage;", can be
  found with: (FIND-SCHEMA 'FOO-PACKAGE::FOO)

* Pathname-based schema lookup no longer ignores file type.

* DO-PROCESS-IMPORT is now responsible for deciding whether or not to
  process an import. It does so by searching for the .proto to be
  imported, and upon finding it, checks to see if we alredy have a
  schema for that pathname. Additionally, it now returns the pathname.

* PROCESS-IMPORTS and PROCESS-IMPORTS-FROM-FILE now defer the work of
  figuring out whether or not an import has been done to
  DO-PROCESS-IMPORT. PROCESS-IMPORTS uses the return value of
  DO-PROCESS-IMPORT to update the PROTO-IMPORTED-SCHEMAS of the schema
  being processed.

New functions:

* ASDF::PROTOBUF-MANGLE-NAME: Takes a pathname, returns a string to be
  used as a filename that represents the original path.
  e.g., #P"/foo/bar/baz.quux" becomes "foo-bar-baz-quux"

* ASDF::PROTOBUF-LISPIZE-PATHNAME: Takes a pathname, returns a new
  pathname with the name mangled as described above, and with a lisp
  type.

So far only CCL appears to consistently handle this.

 - Use 'defparameter' instead of 'defvar' where appropriate.
 - Fix 'defvar' doc strings to distinguish between globals and "thread locals".
 - Avoid using 'nconc'. introduce a new 'appendf' macro instead.
 - Add a comment lamenting the fact that exporting something like 'proto-options'
   also exports the writer '(setf proto-options)'. Fixed in Dylan.

This prevents geodata.proto's FASL from being clobbered by
geodata.lisp's FASL. Before both were geodata.fasl, now the former is
geodata.proto.fasl.

This is generally a good idea.

The client-side stub is now CALL-FOO
The server-side stub is now FOO-IMPL

They are interned in package XXX-RPC, where XXX is the name of the
package being used by the rest of the schema.