btf: don't copy types during CO-RE relocation

[lmb]

btf: add CO-RE benchmark

Signed-off-by: Lorenz Bauer <lmb@isovalent.com>

btf: don't copy types during CO-RE relocation

During CO-RE relocation we need to ignore qualifiers and Typedefs. So far
we've done this by creating a full copy of each target type with those types
stripped out.It turns out that this is really expensive especially for large
types like struct sk_buff.

Stop copying types during CO-RE relocation and instead use the existing
UnderlyingType() and a new helper as() to skip when appropriate. As a rule
of thumb, anytime we do a type assertion during CO-RE relocation we should
use one of these helpers as appropriate.

Embarrassingly this leads to a speedup of 99.99% for large types:

    goos: linux
   goarch: amd64
   pkg: github.com/cilium/ebpf/btf
   cpu: 12th Gen Intel(R) Core(TM) i7-1260P
                               │     base.txt     │               opt.txt   
            │
                               │      sec/op      │    sec/op     vs base   
            │
   CORESkBuff/byte_off-16         2987010.0n ±  2%   311.9n ±  1%   -99.99%
(p=0.002 n=6)
   CORESkBuff/byte_sz-16          3016966.0n ±  0%   348.2n ±  1%   -99.99%
(p=0.002 n=6)
   CORESkBuff/field_exists-16     3036697.5n ±  2%   320.9n ±  2%   -99.99%
(p=0.002 n=6)
   CORESkBuff/signed-16           3133408.5n ±  3%   211.9n ±  1%   -99.99%
(p=0.002 n=6)
   CORESkBuff/lshift_u64-16       3129531.5n ±  5%   358.9n ±  4%   -99.99%
(p=0.002 n=6)
   CORESkBuff/rshift_u64-16       3131090.5n ±  6%   346.9n ±  1%   -99.99%
(p=0.002 n=6)
   CORESkBuff/local_type_id-16        23.83n ±  2%   23.12n ±  2%    -2.98%
(p=0.002 n=6)
   CORESkBuff/target_type_id-16   3660381.0n ± 48%   266.7n ±  4%   -99.99%
(p=0.002 n=6)
   CORESkBuff/type_exists-16      6158062.0n ± 18%   256.6n ±  6%  -100.00%
(p=0.002 n=6)
   CORESkBuff/type_size-16        4724512.0n ±  6%   282.9n ±  1%   -99.99%
(p=0.002 n=6)
   CORESkBuff/enumval_exists-16       703.0n ±  7%   271.2n ± 12%   -61.42%
(p=0.002 n=6)
   CORESkBuff/enumval_value-16        749.0n ±  9%   335.8n ± 18%   -55.16%
(p=0.002 n=6)
   geomean                            319.3µ         240.5n         -99.92%

Signed-off-by: Lorenz Bauer <lmb@isovalent.com>

Updates #1081

[lmb]

FYI @brycekahle