Remove old DDP, FSDP, and DeepSpeed distributed strategies

[w4nderlust]

Summary

• Removes the legacy DDPStrategy, FSDPStrategy, and DeepSpeedStrategy classes along with the DeepSpeedBackend and DataParallelBackend, replacing them with the unified AccelerateStrategy as the sole distributed training backend
• Updates the default strategy from "ddp" to "accelerate" and updates all config validation, tests, and comments accordingly
• HuggingFace Accelerate handles DDP, FSDP, and DeepSpeed via configuration (e.g. fsdp_config, deepspeed_config params) rather than separate class hierarchies

Files deleted

• ludwig/distributed/ddp.py
• ludwig/distributed/fsdp.py
• ludwig/distributed/deepspeed.py
• ludwig/backend/deepspeed.py

Test plan

• Verify local training still works with default (accelerate) strategy
• Verify Ray distributed training works with strategy: accelerate
• Verify LLM finetuning config validation accepts accelerate strategy
• Run integration tests for distributed training

[github-actions]

Test Results

   10 files  +1     10 suites  +1   1h 47m 2s ⏱️ + 1m 9s
3 618 tests +7  3 588 ✅ +5  30 💤 +3  0 ❌  - 1 
3 706 runs  +6  3 664 ✅ +5  42 💤 +2  0 ❌  - 1 

Results for commit 5c83ef7. ± Comparison against base commit 4a10993.

This pull request removes 2 and adds 9 tests. Note that renamed tests count towards both.

tests.integration_tests.test_experiment ‑ test_experiment_model_resume_distributed_gpu[deepspeed]
tests.ludwig.backend.test_ray ‑ test_get_trainer_kwargs[ddp]

tests.ludwig.backend.test_ray ‑ test_get_trainer_kwargs[accelerate]
tests.regression_tests.benchmark.test_model_performance ‑ test_performance[adult_census_income.ecd.yaml]
tests.regression_tests.benchmark.test_model_performance ‑ test_performance[ames_housing.ecd.yaml]
tests.regression_tests.benchmark.test_model_performance ‑ test_performance[mercedes_benz_greener.ecd.yaml]
tests.regression_tests.benchmark.test_model_performance ‑ test_performance[sarcos.ecd.yaml]
tests.regression_tests.model.test_old_models ‑ test_model_loaded_from_old_config_prediction_works
tests.regression_tests.model.test_old_models ‑ test_predict_deprecated_model[respiratory]
tests.regression_tests.model.test_old_models ‑ test_predict_deprecated_model[titanic]
tests.regression_tests.model.test_old_models ‑ test_predict_deprecated_model[twitter_bots]

This pull request removes 1 skipped test and adds 4 skipped tests. Note that renamed tests count towards both.

tests.integration_tests.test_experiment ‑ test_experiment_model_resume_distributed_gpu[deepspeed]

tests.regression_tests.benchmark.test_model_performance ‑ test_performance[adult_census_income.ecd.yaml]
tests.regression_tests.benchmark.test_model_performance ‑ test_performance[ames_housing.ecd.yaml]
tests.regression_tests.benchmark.test_model_performance ‑ test_performance[mercedes_benz_greener.ecd.yaml]
tests.regression_tests.benchmark.test_model_performance ‑ test_performance[sarcos.ecd.yaml]

♻️ This comment has been updated with latest results.