[aws] cache user identity by 'aws configure list'

[aylei]

Cache aws sts get-caller-identity result, indexed by hashing the output of aws configure list. This will speed up sky launch by ~1s in subsequent runs.

original discussion: #3159 (comment)

A simple benchmark (enabled clouds: [aws, gcp, azure, kubernetes, runpod, cudo, paperspace], the code is slightly modified to return before the confirmation is prompted):

# master, avg: 2.558s 
$ /usr/bin/time bash -c 'for i in {1..10}; do SKYPILOT_DISABLE_USAGE_COLLECTION=1 python -m sky.cli launch --gpus H100:8> /dev/null; done;'
       25,58 real        16,57 user        25,29 sys

# PR branch, avg: 1.596s
$ /usr/bin/time bash -c 'for i in {1..10}; do SKYPILOT_DISABLE_USAGE_COLLECTION=1 python -m sky.cli launch --gpus H100:8> /dev/null; done;'
       15,96 real        16,84 user        24,06 sys

# PR branch with https://github.com/skypilot-org/skypilot/pull/4483 picked, avg: 1.321s
$ /usr/bin/time bash -c 'for i in {1..10}; do SKYPILOT_DISABLE_USAGE_COLLECTION=1 python -m sky.cli launch --gpus H100:8> /dev/null; done;'
       13,21 real        16,28 user        11,18 sys

@Michaelvll Please kindly take a look

Tested (run the relevant ones):

• Code formatting: bash format.sh
• Any manual or new tests for this PR (see above benchmark)
• All smoke tests: pytest tests/test_smoke.py
• Relevant individual smoke tests: pytest tests/test_smoke.py::test_fill_in_the_name
• Backward compatibility tests: conda deactivate; bash -i tests/backward_compatibility_tests.sh

[Michaelvll]

Thanks @aylei! This PR looks good to me!

[Michaelvll]

Let's add a docstr for this function for the behavior of caching and returning identity. It might also be good to move the docstr from _sts_get_caller_identity to here.

[aylei]

good catch! just moved the docstring here and added the description of caching behavior

[Michaelvll]

I assume the time tests listed in the PR description has the launch to return immediately after the optimizer has been run, i.e. not going to the confirmation line.

[aylei]

I assume the time tests listed in the PR description has the launch to return immediately after the optimizer has been run, i.e. not going to the confirmation line.

Yes, I updated the PR description to make the result clearer~