MaxOffsetNumber for Table AMs

The notion of TID is based on pages and line pointers, which makes
sense for heapam, but that's not likely to make sense for a custom
table AM.

The obvious answer is to make a simple mapping between a TID and
whatever makes sense to the AM (for the sake of discussion, let's say a
plain row number).

The most natural thing would be to say that we have 48 bits, so it can
just be a 48-bit number. Of course, there are some restrictions on
valid values that complicate this:

* InvalidBlockNumber of 0xFFFFFFFF. Not a problem.
* InvalidOffsetNumber of 0. Not a problem.
* MaxOffsetNumber of 2048. Does this limit really apply to table AMs?
It just seems like it's used when scanning heap or index pages for
stack-allocated arrays. For a table AM it appears to waste 5 bits.
* ginpostinglist.c:itemptr_to_uint64() seems to think 2047 is the max
offset number. Is this a bug?

As a table AM author, I'd like to know what the real limits are so that
I can use whatever bits are available to map from TID to row number and
back, without worrying that something will break in the future. A few
possibilities:

1. Keep MaxOffsetNumber as 2048 and fix itemptr_to_uint64().
2. Change MaxOffsetNumber to 2047. This seems likely to break
extensions that rely on it.
3. Define MaxOffsetNumber as 65536 and introduce a new
MaxItemsPerPage as 2048 for the stack-allocated arrays. We'd still need
to fix itemptr_to_uint64().

Thoughts?

Regards,
Jeff Davis