Streaming CP implementation to handle large datasets - #31

Contributes to #29

This pull request introduces a streaming calibration interface to the conformal calibrators, allowing calibration data to be processed incrementally in small chunks rather than requiring the entire calibration tensor in memory at once. The Ensemble conformal calibrator is extended with a member-streaming API, enabling ensemble members to be processed one at a time, including specialized handling for both quantile and standard deviation modes. Documentation is updated to describe the new streaming APIs and usage patterns.

Streaming Calibration Interface:

• Added a streaming calibration API to the base conformal calibrator, including reset, update, and accumulate_score methods, which allow calibration data to be accumulated incrementally and concatenated only when needed. (src/autouq/calibrators/conformal/conformal.py src/autouq/calibrators/conformal/conformal.pyR41-R119)
• Updated the documentation to explain the new streaming calibration workflow and how it differs from the previous batch-oriented API.

calibrator.reset()
for true_chunk, pred_chunk in calibration_chunks:
    calibrator.update(true_chunk, pred_chunk)
intervals = calibrator.predict(pred_test, alphas=[0.1, 0.2])

Ensemble Calibrator Streaming API:

• Implemented a member-streaming interface for the Ensemble calibrator, with methods such as reset_stream, update_stream, stream_score, and stream_predict, enabling ensemble members to be processed one at a time and supporting both quantile and std modes efficiently. (src/autouq/calibrators/conformal/ensemble.py [1] [2]
• Refactored internal logic to support streaming: ensemble summaries and score computations are now incremental, with Welford’s algorithm for std mode and chunked quantile computation for quantile mode. (src/autouq/calibrators/conformal/ensemble.py [1] [2]
• Updated the ensemble calibrator documentation to describe the streaming member interface and provide usage examples. (docs/ensemble.md docs/ensemble.mdR62-R105)

# Outer loop: one iteration per calibration example (e.g. one forecast
# initialization time). Each example's raw members are only ever held one
# at a time, and are discarded once that example's score has been folded
# into the calibration bank via accumulate_score.
for true_for_this_example, members in calibration_examples:
    calibrator.reset_stream()
    for member in members:                  # one model forward pass each
        calibrator.update_stream(member, true=true_for_this_example)
    score = calibrator.stream_score(chunk_size=2048)   # chunk_size optional
    calibrator.accumulate_score(score)       # feed the base calibration layer

intervals = calibrator.predict(pred_test, alphas=[0.1, 0.2])