libcpp: Optimize Arm search_line_fast

The existing Arm search_line_fast is ancient and not well optimized.
Optimize it for typical inputs (short lines) - what matters is to quickly
get the match result for the first few vectors with minimal initialization
overhead.  We simply loop until a match is found even if the input pointer
is unaligned or close to the end.  Since this may overread, it relies on
*end containing a match and CPP_BUFFER_PADDING >= 16.

Lookup the low 3 bits of each character using 2 64-bit TBLs and compare
the 128-bit result with the original input.  The lookup table contains
the 4 search characters at entries MOD 8, thus if tab[ch % 8] == ch, we
have found a match.  This works because the low 3 bits of the search
characters are unique in ASCII, allowing the use of TBL.  Use a special
value in tab[0] to avoid matching NUL.

The optimized sequence is less than a third of the old code:

	vmov.i8 q12, #7  @ v16qi
	vldr    d22, .L29
.L25:
	vld1.8  {d18-d19}, [r0]!
	vand    q10, q9, q12
	vtbl.8  d16, {d22}, d20
	vtbl.8  d17, {d22}, d21
	vceq.i8 q8, q8, q9
	vaddhn.i16      d16, q8, q8
	vmov    r2, r3, d16     @ int
	orrs    r1, r2, r3
	beq     .L25
	cmp     r2, #0
	rbiteq  r2, r3
	rbitne  r2, r2
	clzeq   r2, r2
	clzne   r2, r2
	addeq   r2, r2, #32
	sub     r2, r2, #61
	add     r0, r0, r2, asr #2
	bx      lr

Performance is significantly better as a result: on a huge trace that replays
the calls from libcpp it is ~24% faster on Neoverse N1.

libcpp/ChangeLog:

	* lex.cc (search_line_fast): Optimized Arm implementation.