Uncap the torchmetrics version, so we can get fixes in 1.0+.

[trevenrawr]

While trying to figure out why my air-gapped installation and deployment of Ludwig was failing to detect my cached copy of the punkt tokenizer for nltk, I came across Lightning-AI/torchmetrics#1779 which got merged into torchmetrics==1.0.0... but Ludwig was avoiding that version. So I forcibly installed torchmetrics==1.2.0 (latest) in my image (after installing Ludwig), ignored pip's words of caution, and managed to run some LoRA-based fine-tuning without a hitch! Since that code path worked okay with the updated library, I'm submitting this PR to remove the pin and leaning on CI to help identify if there are issues on other code paths or if it's truly safe to release.

[justinxzhao]

Thanks @trevenrawr! Let's see what the CI says -- I recall we had to pin originally to mitigate some test failures, though perhaps the test failures are reasonable to fix now.

[github-actions]

Unit Test Results

  6 files  ±0    6 suites  ±0   22m 25s ⏱️ + 1m 41s
12 tests ±0    7 ✔️  -   2    5 💤 +  2  0 ❌ ±0 
60 runs  ±0  30 ✔️  - 12  30 💤 +12  0 ❌ ±0 

Results for commit 248f8d8. ± Comparison against base commit f98b8f6.

This pull request skips 2 tests.

tests.regression_tests.benchmark.test_model_performance ‑ test_performance[ames_housing.gbm.yaml]
tests.regression_tests.benchmark.test_model_performance ‑ test_performance[mercedes_benz_greener.gbm.yaml]

♻️ This comment has been updated with latest results.

[arnavgarg1]

Good news is that there's only 1 test failure! Specifically, in the combinatorial tests: test_number_encoder_loss[config3-dataset3].

I ran this again since I thought it might be transient but seems like it is a persistent error that may require a tiny bit of inspection

[justinxzhao]

Looks like my most recent commit fixed the broken test! I haven't seen anyone make use of RMSE kwargs, so I'm inclined to check this in as is.

Metrics unit tests look good.