feat(air-parser)!: optimize Instruction type layout (#767)

Instead of being boxed for each node, the `Instruction` type contains a boxed variable-size elements.  Thus `Instruction` is quite lean, and the allocator deals with variable-sized elements.

Total number of allocations is more or less same, but less space is wasted for unused memory: previously the Instruction's size was 112 bytes in WASM, now it is 16.

It reduces memory consumption on large AIR scripts (heap size decreased from 7.7MiB to 4.625MiB in parser-10000-100, and for the new parser-calls-10000-100 benchmark, it decreased from 5.115MiB to 4.375MiB).

This is a breaking change as the API changes (though the code that navigates the parsed tree generally should work as is).

air/src/lib.rs

@@ -70,7 +70,7 @@ pub mod parser {
     pub use air_parser::ast::Instruction;
 
     /// Parse an AIR script to AST.
-    pub fn parse(script: &str) -> Result<Box<Instruction<'_>>, String> {
+    pub fn parse(script: &str) -> Result<Instruction<'_>, String> {
         air_parser::parse(script)
     }
 }

air/src/preparation_step/preparation.rs

@@ -68,7 +68,7 @@ pub(crate) fn prepare<'i>(
     run_parameters: RunParameters,
     signature_store: SignatureStore,
 ) -> PreparationResult<PreparationDescriptor<'static, 'i>> {
-    let air: Instruction<'i> = *air_parser::parse(raw_air).map_err(PreparationError::AIRParseError)?;
+    let air: Instruction<'i> = air_parser::parse(raw_air).map_err(PreparationError::AIRParseError)?;
 
     let prev_ingredients = ExecCtxIngredients {
         last_call_request_id: prev_data.last_call_request_id,