Use canonical path to derive name of dataset on Windows

[mjwoods]

If #2084 is merged , netcdf utilities will use Windows paths on mingw-w64. This leads to problems in the output from ncdump, such as the following error from dap4_test/test_meta.sh:

checking: test_anon_dim.2.syn
1c1
< netcdf test_anon_dim.2 {
---
> netcdf .\\results_test_meta\\test_anon_dim.2 {

The default name of the dataset is supposed to be the basename of the input file, minus any file extension. The existing function name_path which determines the default dataset name assumes that paths have Unix directory separators (/), but Windows paths use \\. This PR produces a compatible path format using function NCpathcanonical instead of NCpathcvt.

[DennisHeimbigner]

Good catch. Thanks. This might also be an issue inside the library in nc_create() and
nc_open().

[mjwoods]

I haven't seen any sign of problems in nc_create and nc_open. In my tests, after all of my current PRs (#2106, #2084, #2103, #2104, #2105 ) are merged into master, only 6 tests fail on mingw (compared with 63 for master at commit 09e0e04).

The remaining test failures are:
22 - ncdump_tst_nccopy_w3 (Failed)
24 - ncdump_tst_nccopy_w4 (Failed)
41 - ncdump_test_unicode_directory (Failed)
42 - ncdump_test_unicode_path (Failed)
67 - nc_test_tst_diskless6 (Failed)
71 - nc_test_run_inmemory (Failed)

These failures are related to unicode and mmap support. I plan to look into them later.

[WardF]

Rebasing and retesting to see where it's at.