Count NaN histogram observations in the +Inf bucket

[zhechen-cs]

Histogram.observe(float('nan')) updates the sum but increments no bucket, so both _count and the +Inf bucket omit the observation. Count these observations in the overflow bucket while retaining the NaN sum and leaving finite bucket counts unchanged. This matches the classic-histogram handling described in the Prometheus histogram specification.

Added regressions for ordinary histograms and aggregation from two multiprocess metric files. Both fail on the unmodified base.

Validation on macOS / Python 3.12:

• Targeted histogram and multiprocess tests: 48 passed.
• Full suite: 443 passed, 1 skipped, 28 subtests passed.
• tox -e flake8,isort,mypy and git diff --check: passed.

Local HTTP tests used NO_PROXY=localhost,127.0.0.1,::1 to bypass the local proxy; the proxy's HTTP 502 failure also reproduces on the unmodified base.

@csmarchbanks