This PR overhauls how Babel parser tokenizes the privateIdentifier #name. Currently #name is tokenized as hash and name, in this PR we merge these two tokens into a new privateName token whose value holds the String value of private identifier (without hash). We observe performance gain up to 18%.

$ node --predictable ./benchmark/many-class-private-properties/1-length.bench.mjs
baseline 256 length-1 private properties: 2640 ops/sec ±35.22% (0.379ms)
baseline 512 length-1 private properties: 1700 ops/sec ±1.47% (0.588ms)
baseline 1024 length-1 private properties: 844 ops/sec ±1.24% (1.186ms)
baseline 2048 length-1 private properties: 424 ops/sec ±0.42% (2.359ms)
current 256 length-1 private properties: 3123 ops/sec ±33.93% (0.32ms)
current 512 length-1 private properties: 2010 ops/sec ±0.83% (0.497ms)
current 1024 length-1 private properties: 977 ops/sec ±0.77% (1.023ms)
current 2048 length-1 private properties: 455 ops/sec ±0.85% (2.198ms)

This PR is based on the observation that Babel always does a lookahead when tokenizing #, so we can determine early if an identifier start is following a #, and avoid extra read of the leading identifier character.

By merging # and name we also avoids the tokenizer context update hooks for name-type tokens. We don't need to check of, functions and class for private identifiers anyway.

Since we expose tokens when options.tokens is true, we add a compat routine for tt.privateName which essentially undo the merging, hopefully we can remove it in Babel 8.

Note that Acorn adopts the same approach, which means it is likely that @babel/eslint-parser will have to merge # and name for older Babel versions. By merging tokens in @babel/parser we also do a favour for the @babel/eslint-parser.