table: let a table message hold messages that have no table

[iainmcgin]

Message fields of a table message can now hold messages that have no table: unrolled ones, extern_path types, and well-known types. On whatsapp.proto (752 messages, Box fields, fat LTO), the text section at opt-level = "z" is 1,669,507 bytes unrolled, 1,236,115 with CodecStrategy::Table on #469, and 1,009,611 here; table messages go from 636 to 709 of 752, because 73 of the 116 that fell back on #469 only held a message with a oneof or map, and 43 have one. The guide has the s and 3 figures.

A field's descriptor holds the child's table or a DynVt, function pointers that call the child's Message impl. MsgVt::new_via_message and RepVt::new_via_message are new const fns in buffa::table; a child reached this way decodes from the slice of the nearest enclosing table message, so an overrun fails at that message's end.

Differences from #469:

• A message holding, transitively and through any field kind, a message of the run with a non-default bytes type stays unrolled, because the table decodes a contiguous &[u8] and would copy the Bytes fields. Another crate's child is not inspected, so a holder of google.protobuf.Any (value is Bytes) keeps its table and copies the payload.
• The DynVt write function takes the PreSized cursor that Message::encode uses for every BufMut. Any other sink, including a BufMut passed to Message::write_to, gets each child from a scratch Vec; a Rope copies again.

Instructions per call against all unrolled: a table parent with nine unrolled children is +6.5% to size, +7.3% to encode, +6.5% to decode; a table message of two Timestamp fields is +54%, +41%, +31%. Table-to-table fields cost 0.3% to 1.2% more than on #469.

The unsafe in table/bridge.rs is three thunks that cast *const u8 back to the child type, and the Child methods forwarding to them.

[github-actions]

All contributors have signed the CLA ✍️ ✅
_{Posted by the CLA Assistant Lite bot.}