feat: candid subtype check

[crusso]

Builds on (in reverse order)

• feat: optimized candid subtype check #3170 feat: optimized candid subtype check (adding a global type table of types involved in candid subtype checks)
• feat: candid subtype check, during deserialization #3151 feat: candid subtype check, during deserialization (first integration of subtype checks with deserialization)
• feat: implementation of rts candid subtyping check #3115 feat: implementation of rts candid subtyping check (implementaton of (unextended) type table base candid subtype checks

Since this PR includes more test and fixes to previous PR, its probably best to just review this one as a diff against master:

Over #3170, this PR:

• Introduces a stack allocated memo table allocated on entry to deserialize and shared amongst all recursive calls. The extended boolean flag is replaced by a possibly null memo table (bit_rel_opt) - the table is null for extended deserialization of stable values (for which subtype checks on references are just ommitted since unnecessary).

• Generalises the subtype check and bitrel cache to support sharing across multiple calls and caching of both positive and negative subtype test results:

• Adds 1 bit per pair of types in subtype cache to record true/false outcome separately from plain addition to cache.

• Adjust each exit point that returns false to invalidate the positive subtyping assumption already in the cache (thus recording the negative result in the cache).

Unfortunately caching negative results doubles the space required to store the cache (we now need 2 bits, not 1, for every possible pair of type indices).
This subtyping between type tables with 128 entries each will consume 8K = (2 * 2 * 128 * 128/ 8) of stack space. Not great, but not terrible either.

The first, visited bit (bit 0 per (t1,t2) pair), is stored in unnegated form. The second, related bit (bit 1 per (t1,t2) type pair), is stored in negated form so that assuming the relation holds when first visited is actually a no-op (since the matrix is initialized with zero bits).

Also fixes a lurking bug where I failed to skip the value when the subtype check fails (meaning previous PR were buggy).

• create the cache once on entry to deserialize and share it between calls to idl_sub
• avoid explicitly assuming true by defaulting to full relations on entry (perhaps by inverting true and false bits)
• only produce global type table entries in unextended deserialization. Not worth the effort.
• add basic test for recursive types
• overflow check on Stack.dynamic_alloc
• idl_sub on future types.
• fix broken re-initialization of bitrel on every sub type check
• integrate with changes due to to_candid/from_candid Operations to_candid and from_candid as surface syntax. #3155

[ggreif]

Looks good, but it is a lot.

[crusso]

@chenyan-dfinity can you check the changelog description makes sense?

[crusso]

@chenyan-dfinity PTAL at the changelog.

[chenyan-dfinity]

LGTM