Add soroban-bench-utils, add benchmark tests to measure metering accuracy

[jayz22]

What

Resolves #710
This PR moves the instrumentation helpers from the bench to the host (only turned on for linux and macos), and added a new helper class HostTracker to get the actual resource usage.

Added two test contracts and run them with instrumentation turned on, here are the results comparing modeled vs actual usage:

$RUST_TEST_THREADS=1  cargo test --release --package soroban-env-host --lib -- test::metering_benchmark  --nocapture --ignored

running 2 tests
test test::metering_benchmark::run_add_i32 ... 
model cpu consumed: 1138241, actual cpu insns 1452320 
model mem consumed: 1274466, actual mem bytes 1213168
ok
test test::metering_benchmark::run_complex ... 
model cpu consumed: 1618593, actual cpu insns 1822004 
model mem consumed: 1308496, actual mem bytes 1244696
ok

Which looks decent, but the difference is hard to tell due to VmInstantiation taking up majority of the share.

Why

[TODO: Why this change is being made. Include any context required to understand the why.]

Known limitations

1. The tracking_allocator dependency has been moved from dev to actual. Because the HostTracker is used for "bench" as well (which is under feature testutils, but still needs the dependency). Alternatively we make mod instrumentation a separate package and use it as a dev-dependency. I've moved the tracking-related utils to a separate package soroban-bench-utils to avoid the prod dependency.
2. These tests are disabled by default. For two reasons (added a comment in the test file):
   1. They need to be run with release profile.
   2. They cannot be run with multiple threads due to the gloabal_tracker contention.
3. Only supports os="linux" and "macos", due to the cpu instruction counting library being only available there. The "other" os now has a stub InstructionCounter that returns 0 always.

[graydon]

LGTM, thanks!