Lazy allocation means that we can use a single node block for _all_ empty node blocks. The change in timing for when the definition node is emitted shouldn't affect semantic correctness; the signature is already present for recursive calls.
This change is to make it clearer what is a TODO versus unexpected behavior. I'm doing this now because I feel it's been getting a little confusing in code.
So for example if I write the code a return type `-> i32`, I get the diagnostic output plus the dump of the (invalid) IR:
```
/carbon-lang/toolchain/semantics/testdata/function/basic.carbon:37:10: Semantics TODO: HandleReturnType
cross_reference_irs_size: 1
callables: [
]
integer_literals: [
]
strings: [
]
nodes: [
{kind: CrossReference, arg0: ir0, arg1: block0, type: node0},
{kind: CrossReference, arg0: ir0, arg1: block1, type: node1},
{kind: CrossReference, arg0: ir0, arg1: block2, type: node0},
{kind: CrossReference, arg0: ir0, arg1: block3, type: node0},
]
node_blocks: [
[
],
]
```
For parameters (and in the future, arguments too; generally comma-separated lists) track two node blocks:
1. param_ir: The complete IR.
2. param_refs: Nodes within the IR that are the "root" parameter.
param_refs should allow quick counting of the # of parameters, and more efficient comparison of call args with function parameters. param_ir should be necessary to generate the actual signature.
In order to construct this, this refactors the node_block_stack into its own class, which is reused in params_stack. These carry references to the underlying SmallVector for lazy modification in order to avoid a dependency cycle with SemanticsIR (also see notes on empty node blocks below).
When finalized, the block pair is pushed onto finished_params_stack. That's because node_stack only has space for one thing, and this is two things -- so I'm essentially choosing a trade-off of adding another stack in order to avoid consuming more space in the expectation that most parse nodes have 0 or 1 things to return, and 2 will be very rare.
As factored, this currently consolidates most empty node blocks into a single canonical empty node block. This is because I think empty blocks, i.e. `()`, will be very common. In order to achieve this, SemanticsNodeBlockStack does lazy creation.
An alternative approach would have been to use 1 node block per parameter. We decided against this in order to reduce the number of vectors being created.
Jon Ross-Perkins