Retry transient read errors during input (VCF, binary) streaming in Glimpse phase

[mmorgantaylor]

Fix for input and reference panel streaming, using #212 as a model.

Goal: Make GLIMPSE2_phase resilient to transient read failures when inputs are streamed/localized from a cloud filesystem, by retrying with exponential backoff instead of crashing on the first error — while still failing fast on errors that retrying can't fix.

Two areas in phase changed:

1. Target GL streaming — phase/src/io/genotype_reader.cpp / .h
2. Binary reference panel localization — phase/src/caller/caller_initialise.cpp / caller_header.h

In ligate and concordance, we've added explicit error messages to avoid silent failures, but adding retries would be more complex than I'd like to include in this PR.

Testing:
I ran the same GlimpsePhase task on one reference panel chunk and VCF input, 5000 times in parallel, and checked the crc values for the inputs in the checkpointing file.

• As a positive control, I ran this before applying the changes in this PR, and 9/5000 tasks had differing checksums.
• After these changes, 0/5000 tasks had differing checksums, indicating the issue is resolved

I also confirmed that the output of phase is identical before and after these changes.

[srubinacci]

Thanks for this. I do understand this extends the retry mechanism to VCF files. Should be easy to check and add to main. Is this ready to merge?

[mmorgantaylor]

Is this ready to merge?

@srubinacci I'm testing it now - this actually didn't resolve the issue, but we think we also need to add error handling and retries to the haplotype_set input (in caller_header.h and caller_initialise.cpp)

[mmorgantaylor]

@srubinacci is there any additional testing you'd like to do (or us to do) for this?