Keep the MinGW test objects below the linker's section limit

The MinGW clang job links with a GNU ld that cannot relocate an object
with more than 32767 sections. unit-regression2, unit-regression3 and
unit-msgpack went over it:

- At -O0, the parser and binary reader for parse(), accept() and
  from_*() still emitted the stubs that stand in for recovering, and
  the parser referenced its recovery helpers from code that never runs.
  The stubs are now always inlined, and the parser reaches its helpers
  through the same allow_recovery tag as the stubs, so that this code
  is not even instantiated there. This removes about 300 sections from
  unit-msgpack.
- Every sax_parse call instantiates the parser and binary reader that
  recover from errors. The tests that call it move from
  unit-regression2.cpp (#3989) and unit-regression3.cpp (#5676) into
  unit-sax_parse.cpp, like unit-explicit_instantiation.cpp got a file
  of its own (#5511), and CONTRIBUTING.md points there.

Section counts of the MinGW Debug objects (clang 19, -g0), develop / PR:
unit-msgpack 32281 / 32558, unit-regression2 31272 / 31278,
unit-regression3 32392 / 27462, unit-sax_parse - / 19836.

Signed-off-by: Niels Lohmann <mail@nlohmann.me>

.github/CONTRIBUTING.md

@@ -111,6 +111,8 @@ context which existing file needs to be extended, and only very few cases requir
 When fixing a bug, edit `unit-regression3.cpp` and add a section referencing the fixed issue.
 `unit-regression2.cpp` holds the older tests; the two files exist because a single one grew large enough for the
 MinGW linker to fail relocating it, so please keep adding to the smaller file rather than growing the larger one.
+Regression tests that call `sax_parse` go into `unit-sax_parse.cpp` instead: every call instantiates the parser and
+the binary reader that recover from errors, which grows a test file considerably.
 
 #### Exceptions