fix(cranelift): publish the caller's rwasm counters before a tail call

[dmitry123]

Summary

• The rwasm stack-limit emulation (feat(cranelift): emulate the rwasm stack limits in compiled code #13) keeps one packed counter (call depth, frame base) in the store. A plain call publishes the callee's values and nothing restores the caller's afterwards; a tail call published nothing at all. So a frame that made a plain call and then a return_call handed the tail callee the callee's counters, and every iteration of a tail-recursive loop with a call in its body counted as one more frame until StackOverflow at N_MAX_RECURSION_DEPTH. The rwasm VM runs the same module, so the two backends diverged.
• return_call, return_call_indirect and return_call_ref now store the caller's own counters (the SSA value its prologue loaded) before the call, like ReturnCallInternal. One extra store per tail-call site; modules without tail calls compile to the same code.
• Version 45.0.0-rwasm.4 for the three published crates.

Reproduction

rwasm tests/stack-overflow.rs, new case tail_calls_after_a_call_do_not_count_towards_the_recursion_limit (return_call and return_call_indirect after a call, 100 × N_MAX_RECURSION_DEPTH iterations):

• rwasm.3 from crates.io: [("rwasm", Ok(I32(0))), ("wasmtime", Err(StackOverflow))]
• this branch: results follow in a comment once the local run completes (draft until then).

Impact

Reachable only by modules with tail calls after a non-tail call in the same function. None of the Fluent networks' historical system-runtime hints use return_call, so nothing on chain is affected; the fix is needed before any tail-call payload lands.

Companion: rwasm fix: pin wasmtime-rwasm 45.0.0-rwasm.4 and cover tail calls after a call (test, docs, changelog).

[dmitry123]

Local run of rwasm tests/stack-overflow.rs (rwasm v0.7.1 + the test from fluentlabs-xyz/rwasm#224) with the three crates patched to this branch:

• tail_calls_after_a_call_do_not_count_towards_the_recursion_limit: rwasm.3 → [("rwasm", Ok(I32(0))), ("wasmtime", Err(StackOverflow))]; this branch → both Ok(I32(0)) for return_call and return_call_indirect at 100 × N_MAX_RECURSION_DEPTH iterations.
• Whole stack-overflow suite: 16 passed. fuel_alignment 43, wide_arithmetic 4, execution_contracts 21, instance_replacement 16, fluentbase 3 passed; lib: test result: ok. 154 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 41.08s ##### DONE (07:59:15)

Ready for review. After merge: tag v45.0.0-rwasm.4 so publish.yml publishes the three crates, then rwasm#224 moves its pin.

[dmitry123]

Added a fork-local regression test, tests/all/rwasm_stack.rs (cargo test --test all rwasm_stack): it sets Config::rwasm_stack_limits directly, appends the rwasm.frames section itself, and covers two cases.

• plain_recursion_traps_at_the_depth_limit (control): max_call_depth - 1 nested calls run, one more traps StackOverflow.
• tail_call_after_a_call_keeps_the_caller_frame: return_call and return_call_indirect after a plain call, 100 × max_call_depth iterations.

lowering	control	tail call after a call
func_environ.rs from rwasm.3 (d0ea7f7)	ok	FAILED, StackOverflow
this branch	ok	ok

[hedwig0x]

Reviewed the compiler lowering, regression coverage, and crate version updates. The focused rwasm stack tests pass locally; no blocking issues found.