btf: split string table ahead of time

[lmb]

We currently read the string table into one big byte slice, and
allocate substrings in Lookup. This forces us to hang on to one
large allocation for the string table, and many small allocations
for substrings. Calling Lookup for the same offset also allocates
on each invocation.

Instead, split the string table into substrings at load time. In
Lookup we can then use a binary search to find the correct string.
As a result, both overall memory and allocations drop:

name            old time/op    new time/op    delta
ParseVmlinux-4    76.6ms ±12%    73.9ms ± 4%     ~     (p=0.343 n=4+4)

name            old alloc/op   new alloc/op   delta
ParseVmlinux-4    92.9MB ± 0%    84.3MB ± 0%   -9.22%  (p=0.029 n=4+4)

name            old allocs/op  new allocs/op  delta
ParseVmlinux-4      786k ± 0%      693k ± 0%  -11.87%  (p=0.029 n=4+4)

To keep the speed roughly the same I had to "vendor" sort.Search and
specialised it for uint32, since the function was prominent in profiles.

[lmb]

cc @paulcacheux, this was the idea I mentioned, please take a look.

[lmb]

I also tried replacing offsets, strings with a map[uint32]string but its slower:

name            old time/op    new time/op    delta
ParseVmlinux-4    74.3ms ± 5%    82.4ms ± 3%  +10.87%  (p=0.029 n=4+4)

name            old alloc/op   new alloc/op   delta
ParseVmlinux-4    84.3MB ± 0%    87.8MB ± 0%   +4.16%  (p=0.029 n=4+4)

name            old allocs/op  new allocs/op  delta
ParseVmlinux-4      693k ± 0%      693k ± 0%     ~     (p=0.943 n=4+4)

[paulcacheux]

This looks great, thanks for working on this