[SOL] Expand memcmp earlier (take 2)

[LucasSte]

Problem

It has been pointed out that comparing pub keys with the equal operator consumes more CUs than it should. I found the problem to be that LLVM lowers memcmp too late in the pipeline stage, so that it still contains suboptimal code when going to machine code generation.

Solution

Lower memcmp earlier.

Benchamarks

Programs/sbf

  SBF program                          expected actual  diff
  solana_sbf_rust_128bit                    770   770    +0 ( +0%)
  solana_sbf_rust_alloc                    4767  4767    +0 ( +0%)
  solana_sbf_rust_custom_heap               324   324    +0 ( +0%)
  solana_sbf_rust_dep_crate                  20    20    +0 ( +0%)
  solana_sbf_rust_iter                     1513  1513    +0 ( +0%)
  solana_sbf_rust_many_args                1281  1281    +0 ( +0%)
  solana_sbf_rust_mem                      1229  1195   -34 ( -3%)
  solana_sbf_rust_membuiltins               277    69  -208 (-75%)
  solana_sbf_rust_noop                      342   342    +0 ( +0%)
  solana_sbf_rust_param_passing             108   108    +0 ( +0%)
  solana_sbf_rust_rand                      301   301    +0 ( +0%)
  solana_sbf_rust_sanity                  13578 13576    -2 ( -0%)
  solana_sbf_rust_secp256k1_recover       88376 88376    +0 ( +0%)
  solana_sbf_rust_sha                     21984 21984    +0 ( +0%)

No change to solana or Pinocchio entrypoint

Pubkey comparison

Name	CUs	Delta
Different	10	-2
Same	17	-3

P-token (after reverting solana-program/token#90 and removing pubkey_eq from Pinocchio).
The comparison is against LLVM master using p-token unmodified and this PR using p-token without pubkey_eq.

Instruction	Master	This PR	Diff
initialize_mint	106	108	+2
initialize_account	159	157	-2
initialize_multisig	211	221	+10
transfer	76	76	0
approve	129	136	+5
revoke	102	102	0
set_authority	131	130	-1
mint_to	124	123	-1
burn	129	129	0
close_account	121	121	0
freeze_account	150	150	0
thaw_account	146	146	0
transfer_checked	105	105	0
approve_checked	169	183	+14
mint_to_checked	173	177	+4
burn_checked	141	147	+6
initialize_account2	176	179	+3
initialize_account3	246	247	+1
initialize_multisig2	342	342	0
initialize_mint2	228	229	+1
amount_to_ui_amount	423	424	+1
ui_amount_to_amount	703	706	+3
initialize_immutable_owner	38	38	0
sync_native	62	62	0
batch	1499	1496	-3

[LucasSte]

Relative to #170, the numbers for programs/sbf seem better and for p-token only slightly better (@febo agreed with that) . Since P-token numbers fluctuate a lot, I ran an Anchor tic-tac-toe game and compared the numbers to ensure this PR is not a regression.

PS: I am not repeating all the numbers for 'play' since it is a call to the same function with a different board configuration. The whole list for master is here and for this PR is here.

Instruction	Master	This PR	Diff
SetupGame	5966	5937	-29
Play - 1	2480	2436	-44
Play - 2	2539	2493	-46
Play - Last	3105	3044	-61