read_nsrdb_psm4: parse header with the csv module to keep quoted commas

[gaoflow]

• Closes Issue with importing NSRDB spectral-on-demand files with pvlib.iotools.read_nsrdb_psm4 #2736
• I am familiar with the contributing guidelines
• Tests added
• Updates entries in docs/sphinx/source/reference for API changes. (n/a — no API change)
• Adds description and name entries in the appropriate "what's new" file in docs/sphinx/source/whatsnew for all changes.
• New code is fully documented. (behavior unchanged for existing files; no public-API docstring change)
• Pull request is nearly complete and ready for detailed review.

What this fixes

read_nsrdb_psm4 parsed its three header lines with a naive str.split(','):

metadata_fields = fbuf.readline().split(',')
metadata_values = fbuf.readline().split(',')
columns        = fbuf.readline().split(',')

The NSRDB spectral-on-demand CSVs reported in #2736 have quoted column
names that contain commas, e.g.

..., "GaAs (Bauhuis et al., 2009)","InGaP (Gray, 2008)", ...

These are valid CSV (the commas are inside quotes), and pandas.read_csv
parses the data rows correctly — but str.split(',') splits each quoted name
into multiple fragments, inflating the column count. The mismatch between the
mis-split names/usecols and the correctly-parsed data then raises on read.

The change

Parse the three header lines with the csv module (which honors quoting)
instead of str.split(','). For ordinary (unquoted) files this is identical
to the previous behavior, so the existing readers are unaffected.

This addresses the parsing crash that @kandersolar confirmed should be
supported. The further map_variables=True unit handling for spectral files
(W/m²/µm → W/m²/nm) mentioned in the issue is a separate enhancement and is
left out of scope here.

Reproduction (before this PR)

from io import StringIO
from pvlib.iotools import psm4

content = (
    "Source,Location ID,City,State,Country,Latitude,Longitude,Time Zone,"
    "Elevation,Local Time Zone,Version\n"
    "NSRDB,1,-,-,-,40.0,-105.0,-7,1600,-7,4.0.1\n"
    'Year,Month,Day,Hour,Minute,GHI,"GaAs (Bauhuis et al., 2009)",'
    '"InGaP (Gray, 2008)"\n'
    "2023,1,1,0,0,0,0.1,0.2\n"
    "2023,1,1,1,0,5,0.3,0.4\n"
)
psm4.read_nsrdb_psm4(StringIO(content), map_variables=False)
# ParserError: Too many columns specified: expected 10 and found 8

After the fix the quoted columns survive intact
('GaAs (Bauhuis et al., 2009)', 'InGaP (Gray, 2008)').

A regression test (test_read_nsrdb_psm4_quoted_columns_with_commas) is added
that fails on main and passes with this change; the existing
read_nsrdb_psm4 tests continue to pass.

[kandersolar]

Nice PR, thanks @gaoflow