Make sinks, sources, and pipelines generic on their errors

[passcod]

Implementation of #29.

• PipelineError becomes generic: PipelineError<SrcErr: SourceError, SnkErr: SinkError>
• We introduce the SourceError and SinkError marker traits which the respective errors must implement. This is a slight workaround because PipelineError<SrcErr: Error, SnkErr: Error> doesn't work (I tried, the type system doesn't like it), but it doesn't harm ergonomics too much.
• We add associated Error types on the Sink, BatchSink, and Source traits, with bounds on the respective SinkError and SourceError traits.
• We add a CommonSourceError type which encodes the common error variants implemented in the pipeline/source glue code, rather than the ones defined by the individual source/sink types.
• For convenience we add two never types: InfallibleSourceError and InfallibleSinkError and implement the SourceError and SinkError traits on them. This saves making your own useless error types when implementing sinks and sources that don't error.
• Finally we move the DuckDB and BigQuery errors to their own types, implement the traits, and wire everything up correctly.

Something remarkable with this (which even surprised me when I wrote it!) is that while this is a breaking API change, it doesn't require any change of the typical downstream code, as illustrated by this PR not changing anything in the replicator crate. Type inference takes care of it.

[imor]

Hey @passcod thanks for the PR. I haven't been able to find time in the last few days but I think I'll have some time tomorrow to review.

[imor]

The PR looks great, thanks for the effort @passcod. The only question I had was about explicit map_errs added, could these have been avoided? In any case I'm merging it now and will see if these map_errs can be removed later or not.

[imor]

Oh I see why the #[from] annotations were changed and map_err were used, since both of them are Errors they conflict. I think this is also somehow related to your second point about the type system not liking using traits directly in the bounds. No worries, everything's good.