fix(dlt): compare dlt version semantically for destination client

[anxkhn]

The dlt destination-client selection in sqlmesh/integrations/dlt.py gated the
1.10.0 API change with a string comparison:

if dlt.__version__ >= "1.10.0":
    client = pipeline.destination_client()
else:
    client = pipeline._sql_job_client(schema)  # type: ignore

String comparison is lexicographic, not semantic, so it is wrong for any version
with a double-digit minor. For example "1.9.0" >= "1.10.0" evaluates to True
(the character '9' sorts after '1'), and likewise for "1.2.0" and
"1.5.10". The result is that every 1.x version takes the >= 1.10.0 branch,
and the pre-1.10 _sql_job_client fallback (added in #4223 for compatibility
with older dlt) is unreachable.

This is currently latent rather than crashing, because dlt versions below 1.10.0
happen to expose destination_client() as well. But the version guard is
silently incorrect and the intended fallback never runs.

The fix uses packaging.version.parse for a semantic comparison, mirroring the
idiom already used in sqlmesh/core/config/connection.py
(version.parse(...) < version.parse("...")). packaging is already a
top-level dependency, so no new dependency is introduced.

The client selection is extracted into a small _get_destination_client helper
so the version gate can be unit tested without a live dlt pipeline.

Test Plan

• Added test_get_destination_client_version_gate in
  tests/integrations/test_dlt.py: a parametrized, pure-Python test that
  monkeypatches dlt.__version__ and asserts 1.9.0, 1.2.0, 1.5.10 select
  _sql_job_client while 1.10.0, 1.28.1, 2.0.0 select
  destination_client. It fails under the old string comparison for the pre-1.10
  cases and passes after the fix. No warehouse or live dlt pipeline required.
• pytest tests/integrations/test_dlt.py -> 7 passed.
• make style (ruff check, ruff format, mypy) clean.

Checklist

• I have run make style and fixed any issues
• I have added tests for my changes (if applicable)
• All existing tests pass (make fast-test)
• My commits are signed off (git commit -s) per the DCO

[StuffbyYuki]

Thanks for this!

Given destination_client() exists on all realistically supported dlt versions and is what everyone already hits due to the string-compare bug, would you consider simplifying to always use destination_client() (or pinning dlt>=1.10) instead of maintaining a version gate?

Let me know what you think