Perf improvements

[EdMcBane]

What kind of change does this PR introduce?

A new iter_scan method to reuse a borrowed Row; a few additional performance improvements.

What is the current behavior?

ForeignDataWrapper::iter_scan requires the implementor to build a Row, which is quite expensive.

What is the new behavior?

ForeignDataWrapper::iter_scan_borrowed takes a mutable ref to a Row owned by the caller; the implementor just needs to populate it.
A default implementation is provided that simply delegates to ForeignDataWrapper::iter_scan and swaps the returned Row.

Additional context

On my machine (intel 8700k), in release mode, pg15, with a sample FDW that yields 10M "helloworld" rows, the time taken to execute SELECT * FROM bench WHERE id < 0 are:

• Original: 3244,732 ms (00:03,245)
• Without debug2! on iter_scan: 1565,417 ms (00:01,565)
• Reusing Row: 1241,977 ms (00:01,242)

I've been unable to run the tests on this, the docs seem out of date and cargo test fails to compile.

[EdMcBane]

Example wrapper used for "benchmarks"

https://github.com/EdMcBane/wrappers_bench

[burmecia]

That's great improvement, thanks @EdMcBane ! I think this should be the default interface and we will make it in the next release.