feature: fast initializer for conda environments

[saikonen]

adds a feature flag to switch between regular micromamba and a custom binary for bootstrapping of environments when executing remotely.

also adds deferring of micromamba install until the binary location is actually requested.

some benchmarking results with @pypi

@pypi(python="3.10", packages={"torch": "2.5.1"})

Time taken for fast_setup_environment: 11.00 seconds
Time taken for setup_environment: 28.31 seconds

@pypi(python="3.10", packages={"matplotlib": "3.9.2"})

Time taken for fast_setup_environment: 5.29 seconds
Time taken for setup_environment: 13.53 seconds

@pypi(python="3.11", packages={
    "seaborn": "0.12",
    "fastapi": "0.103.1",
    "scikit-learn": "1.6.1",
    }
)

Time taken for fast_setup_environment: 6.02 seconds
Time taken for setup_environment: 10.04 seconds

[savingoyal]

can you remove this config and introduce the path to actual binary in the install_fast_initializer method instead? also, if you can note some latency improvements in the PR, that will be great!

[savingoyal]

looks good! can you verify that this works with dev-gcp and dev-azure too before merging?