This commit reworks the handling of `zig fmt: off/on`. A motivating
example is shown below:
const c = d;
// zig fmt: off
// comment
const a = b;
// zig fmt: on
Before processing the decl `const a = b;`, `renderRoot` looks for
`zig fmt: off` that appears between this decl and the previous one. If
it finds one, it searches for the next `zig fmt: on` that re-enables
reformatting (or EOF if none was found), and copies the input code
between `zig fmt: off` and `zig fmt: on` to the output stream. After
that, it proceeds to the next decl.
The important thing to notice here is that `renderTopLevelDecl` emits
line comment tokens that follow the decl. Therefore, when copying code,
we must be careful not to copy the line comment tokens that already have
been written to the output stream. The original code failed to take this
fact into consideration. It did skip `// zig fmt: off`, but not the
remaining ones. As a result, when the above example is fed as input, it
duplicated the line `// comment`.
The `arena` instance being used bythe parse tree was valid and
pointed to valid memory, but existed as a local variable inside the
stack frame of the `parse` function (the `const arena`), which was never
stored anywhere before leaving the scope.
This meant that code above the `parse` function saw a valid instance of
an `ArenaAllocator` that pointed to the same backing memory, but didn't
posess any of the local state built up after the call to `parseRoot`,
basically the caller saw an empty arena.
This meant that when `deinit` was called, it saw an Arena with 0
allocations in it's `buffer_list` and wasn't able to destroy any of the
memory. This caused it to leak and caused FailingAllocator to balk.
The fix is to make sure the parse tree is using the same instance of
ArenaAllocator as is reported up the call stack, the one inside the
`Tree{}` object. I'm not sure why that field is marked with a comment
to remove it, as it's used by the `std.ast.Tree.deinit()` function, but
this change seems to solve the problem.
