Optimize CEL PrattParser and Lexer performance

[copybara-service]

Optimize CEL PrattParser and Lexer performance

Improve parsing throughput and reduce memory allocations across CEL expressions.

Measured with CelParserBenchmark (parseOnly, built -c opt), comparing three
parsers back to back in one session: ANTLR, the Pratt parser before this change,
and the Pratt parser after it.

Objects allocated per parse:

Case	ANTLR	Pratt before	Pratt after	Pratt vs ANTLR	Delta this CL
SMOKE_TEST	357	131	123	2.9x smaller	-6.1%
CHAINED_ORS	968	374	350	2.8x smaller	-6.4%
LIST_COMPREHENSION	512	218	166	3.1x smaller	-23.9%
MESSAGE_CREATION	1,253	502	427	2.9x smaller	-14.9%
LONG_LIST	81,794	19,310	19,271	4.2x smaller	-0.2%

Bytes allocated per parse:

Case	ANTLR	Pratt before	Pratt after	Pratt vs ANTLR	Delta this CL
SMOKE_TEST	12,256	3,776	3,608	3.4x smaller	-4.4%
CHAINED_ORS	32,160	10,288	9,864	3.3x smaller	-4.1%
LIST_COMPREHENSION	17,320	6,120	4,736	3.7x smaller	-22.6%
MESSAGE_CREATION	43,128	13,888	12,240	3.5x smaller	-11.9%
LONG_LIST	2,907,488	553,160	568,080	5.1x smaller	+2.7%

Wall clock, mean of 3 caliper trial medians:

Case	ANTLR	Pratt before	Pratt after	Pratt vs ANTLR	Delta this CL
SMOKE_TEST	4,940 ns	733 ns	692 ns	7.1x faster	-5.7%
CHAINED_ORS	14,641 ns	2,090 ns	2,041 ns	7.2x faster	-2.3%
LIST_COMPREHENSION	7,514 ns	1,640 ns	1,218 ns	6.2x faster	-25.7%
MESSAGE_CREATION	20,979 ns	3,962 ns	3,487 ns	6.0x faster	-12.0%
LONG_LIST	1,616,631 ns	144,980 ns	142,583 ns	11.3x faster	-1.7%

LONG_LIST is the one case that allocates slightly more than before. It is an
extreme outlier (1,000 list elements, ~20k objects per parse) and the +2.7% comes
from letting the positions map grow from its default capacity instead of
presizing it; presizing cost more on every other case, so the tradeoff is worth
it. The map is removed entirely later in this series.