experiment: lazy migration for stable memory for regions

[crusso]

builds on the unwieldly #3768.

Introduces lazy migration of stable memory format 0/1 (no ESM, some ESM) to 2 (Regions) by

1. Adding an explicit global (StableMem.get/set_version) that tracks the current version.
2. Transitioning from 0 to 1 on first (acutally growing) StableMem.grow.
3. Transitioning from 0/1 to 2 on first Region.new().
4. Testing current version on each (legacy) StableMemory operation and dispatching to the correct legacy/region implementation.

As a result programs have zero overhead until they allocate the first region and all programs can access Region.mo without having to set a flag.

In addition, ExperimentalStableMemory retains the same low space overhead as before, at least until the first region is allocated. Each ESM read-write is slightly slower since it has to switch on the version, but I can live with that.

Avoids:

• the need to set a flag to enable Regions (and the need to explain and document that)
• significant overhead of region metadata (1-8MB) until regions are actually used.
• language dependence on compiler flag (c.f. Haskell).

Removes the need for a user facing --stable-regions flag. Instead, for testing, the flag now just forces eager
(not lazy) migration to version 2.

Drive-by: remove prim stableMemoryRegion since its dangerous (the region returned won't alias the internal region0 after an upgrade without more work).

[github-actions]

Comparing from ee79edf to 36a5eda:
In terms of gas, 1 tests regressed, 2 tests improved and the mean change is +6.6%.
In terms of size, 5 tests improved and the mean change is -0.2%.

[matthewhammer]

LGTM, modulo the StableMem and StableMemory distinction in compile.ml. Maybe we can combine and consolidate these in the future, or explain why they are separate?

[matthewhammer]

Each ESM read-write is slightly slower since it has to switch on the version, but I can live with that.

Okay. Sounds good.

Removes the need for a user facing --stable-regions flag. Instead, for testing, the flag now just forces eager
(not lazy) migration to version 2.

Also sounds good.

[crusso]

@matthewhammer thanks for taking a look!

[luc-blaeser]

Just out of curiosity: Is wasm-opt failing or creating invalid code?

[crusso]

I don't fully recall, but I think it was failing.

[crusso]

Commit 280ed54 produces this error. Basically I was trying to conditionally call one of two functions with args already on the stack and wasm-opt barfed:

test-bench> @@ -0,0 +1,4 @@
test-bench> +Error: Failed to read module
test-bench> +
test-bench> +Caused by:
test-bench> +    [parse exception: Block requires more values than are available (at 0:57416)]

[crusso]

I should probably be a good citizen and report the bug, but I'd rather make progress..

[luc-blaeser]

Great simplification with lazy region metadata allocation.
Also, very nice refactoring regarding region0.