Improve hfile_s3.c error handling

[jmarshall]

The rewritten S3 backend does not translate libcurl error values into errno values; in some circumstances it returns an error indication but does not set errno at all, leaving it set to whatever it was set to previously. So for example a malformatted URL results in a spurious ENOENT error message:

$ htsfile s3:expect_invalid_argument
htsfile: can't open "s3:expect_invalid_argument": No such file or directory

This led to a bioconda package build test failure (bioconda/bioconda-recipes#64884):

running run_test.py
    pysam.AlignmentFile('s3:expect_invalid_argument')
  File "pysam/libcalignmentfile.pyx", line 752, in pysam.libcalignmentfile.AlignmentFile.__cinit__
  File "pysam/libcalignmentfile.pyx", line 952, in pysam.libcalignmentfile.AlignmentFile._open
FileNotFoundError: [Errno 2] Could not open alignment file: No such file or directory: 's3:expect_invalid_argument'

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "/opt/conda/conda-bld/pysam_1777293322095/test_tmp/run_test.py", line 15, in <module>
    assert e.errno == errno.EINVAL
AssertionError

because the test was written to validate the expected particular errno value that the libcurl backend produces. This particular test has since been rewritten in a more robust way, but nonetheless hfile_s3.c should enable meaningful error messages.

Also requests for buckets which you don't have permission to access or non-existent buckets produce unhelpful errors:

$ htsfile s3://bucket/foo.sam
htsfile: can't open "s3://bucket/foo.sam": Resource temporarily unavailable
$ htsfile s3://bork-asdfghjk/foo.sam  # Nonexistent at the time of writing
htsfile: can't open "s3://bork-asdfghjk/foo.sam": Resource temporarily unavailable

With -vvvv you can see that errno is being set appropriately from the 403 or 404 response, but it is overwritten with a spurious EAGAIN errno value before it is returned to the caller.

Hence this PR:

Libcurl API functions mostly do not set errno, so set it explicitly based on libcurl API error codes. Copy easy_errno() from hfile_libcurl.c, alongside the previously copied http_status_errno().

Also preserve errno in s3_read_open()/s3_write_open()'s error cleanup. Otherwise it will often be changed to EAGAIN by curl_easy_cleanup(), which is called during cleanup_local().

[jmarshall]

curl_global_init() always returns CURLE_FAILED_INIT on error, so simply translate that never-happens into a suitably catastrophic errno value.

[jmarshall]

In the unlikely event curl_easy_getinfo() fails, response_code will usually not have been written to. So separating these cases avoids reading an uninitialised variable in http_status_errno(response_code) below.

[whitwham]

Thank you.