adds a vcsim based test suite and runs it in CI

[semx]

Following up on #474: the test setup, before the remaining PRs.

The VMware source could only be exercised against a live vCenter, which is why a bug like #476 reached users instead of a test run. vcsim, the simulator from the govmomi project, replays an inventory captured with govc object.save, so the real pyVmomi code path runs in CI with no vCenter and no NetBox involved.

What is in here

• tests/fixtures/vcsim/ holds the two captures @ionoci contributed in Maintainer wanted - sunsetting this repository by 30.11.2026 #474 (a 6.7.0 and an 8.0.3 vCenter). The vcsim backed tests are parametrized over every archive in that directory, so dropping in a new capture covers it without touching test code. The README there documents how to take one.
• tests/conftest.py starts one vcsim per capture on a free port, gives each test a clean in-memory NetBoxInventory, and writes the source a generated settings.ini. It also opens a second, independent pyVmomi connection to the same simulator, so the tests compare the sync against the SDK rather than against themselves.
• tests/test_vmware_source.py asserts that every VM and host is synced; host serial, model and manufacturer match the hardware (allowing for manufacturer normalisation, Dell Inc. to Dell); VM cpu and memory match the hypervisor; interface MAC addresses match; guest IPs of VMs with running tools are synced and bound to an interface; two runs produce the same inventory; and a run against objects that already exist in NetBox creates nothing new.
• .github/workflows/test.yml installs vcsim and runs pytest on pushes and pull requests to main and development.

Without vcsim installed the vcsim backed tests skip rather than fail, so pytest still works on a machine that only has the Python dependencies.

Verification

16 tests pass on Python 3.13 (the version the workflow uses) and on 3.14.

To show the suite actually catches regressions rather than just passing, I injected four faults into the source, one at a time, and ran it against the 6.7.0 capture:

injected fault	result
memory reported in GB instead of MB	1 failed, 7 passed
guest IPs never collected (the shape of #476)	1 failed, 7 passed
a VM silently dropped during parsing	4 failed, 4 passed
MAC address dropped from interface data	3 failed, 5 passed

After reverting each one the suite was green again.

Notes

• The captures are about 1.2 MB each and contain host names, VM names, MAC addresses and guest IPs of the source environments. They were posted publicly in Maintainer wanted - sunsetting this repository by 30.11.2026 #474, so this only moves them into the repository; say the word if you would rather keep them out of git and fetch them in CI instead.
• While writing the MAC assertions I noticed that one VM in the 6.7.0 capture is parsed three times in a single run, which leaves duplicate unsaved NBMACAddress objects for its interfaces (82 objects for 64 distinct interface and address pairs). It does not survive a rerun against saved objects, so it looks like extra POSTs rather than lasting damage. The test compares address sets so it does not fail on it. I will open a separate issue with the details.