compiler: keep the C++ throw site out of user diagnostics (#557)

[ohohoreilly]

Closes #557.

Before

$ orlyc -o . bad.orly
2:28-2:47 [orly/type/add_visitor.h, 50]This expression is invalid.

After

$ orlyc -o . bad.orly
2:28-2:47 The left side of a mutation must be a stored value, but this is an address. Name the value stored at it with `*` and its type, as in `*<['counter']>::(int) += n`.

What changed

• TSourceError keeps the throw site as a field (GetCodeLocation()) instead of splicing it into what(). Every one of the ~2000 throw sites already passed the location as its own argument, so only error.h and the two print sites change.
  • orlyc --compiler-locations appends it. --debug/-d is taken, so this has its own flag.
  • WS source_error replies carry it in a separate compiler_loc key, so result is a clean message. This is documented in docs/PROTOCOL.md.
  • TImpossibleError keeps it in the message, now with a space after it. For an internal compiler error the location is the report.
• The target of a mutation must now be a stored value. TMutate::TypeCheck checks this before running the operator's visitor:
  • A bare address gets the message above.
  • A plain value (x += n with x = 1) used to segfault orlyc with no output. On the released ghcr.io/orlyatomics/orly:latest it exits with code 139 and prints nothing. The crash is in TMutation::GetValType, which does As<TMutable>() on the target's type and dereferences the result unchecked. It now gets a clean error.
  • Sequences are excluded so they keep their existing diagnostic.

Tests

• Two new diagnostic pins, diag_mutate_address and diag_mutate_plain_value, added to .xfail under Compiler diagnostics leak internal C++ source paths to users ([orly/type/add_visitor.h, 50]) #557 like the synth/expr/symbol: improve compiler diagnostic messages #314 pins.
• Two existing baselines quoted a compiler path in a user-facing line, and that path is removed: diag_assign_type_mismatch and sequence_in_effecting. The two compile failure[orly/compiler.cc] lines come from a different error class that only prints under -d, so they are unchanged.
• The new baselines were written by hand, not regenerated, so lang_test in CI is the real check on their positions.