Replace ANTLR with hand-rolled parser

[stackoverflow]

Fixes #906
Fixes #888
Fixes #927
Fixes #931
Fixes #932

Incompatible changes:

• Object entries in the same line now require a semicolon (properties and elements are unaffected):

// not valid anymore
foo { ["1"] = 1 ["2"] = 2 }

// requires ;
foo { ["1"] = 1; ["2"] = 2 }

Reason: not requiring semicolons or newlines requires too much backtracking in the parser, degrading performance. Also both our IntelliJ plugin and the LSP don't allow this syntax and show an error.

• Some invalid string escapes are not accepted anymore (see Parser accepts wrong string escapes #888)
• Unseparated object members require a space, semicolon or newline (see Spread elements inside an object body don't need separators #932)

Our current ANTLR parser is quite slow, and, depending on the codebase, can be the slowest part of running Pkl.

Some results from parsing all snippet tests in pkl-core showing ~100x performance improvement:
ANTLR elapsed: 7122ms (7.1s)
New parser elapsed: 78.73ms

This is still a draft, many tests are still failing and repl parsing is still using the old parser.

[KushalP]

I like this comment, but it would be useful to link through to the method where this behaviour is expected in the parser.

[bioball]

Did a first pass! Still have much more to review.

[bioball]

Should we make this allocation free?

One option is to just make children a field on the class.

Alternatively, maybe this should instead be visitChildren. In that case, this becomes:

<T> @Nullable T visitChildren(ParserVisitor<T> visitor) {
  for (var parameter : parameters) {
    parameter.accept(visitor);
  }
  for (var member : members) {
    member.accept(visitor);
  }
}

[stackoverflow]

I kept the children() method there but refactored the visitor not call it anymore. So no allocations are being made in our visitors.