[PyTorch] Add tuple inline storage (pytorch#64066)

Summary:
Pull Request resolved: pytorch#64066

I noticed a bunch of time being spent heap-allocating Tuples
in the unpickler. 1-, 2-, and 3-element Tuples are apparently common
enough that they get their own bytecode instructions, so I decided to
try also giving them their own representation. We store up to 3
IValues inline in `Tuple` rather than doing a second heap allocation
for a `std::vector<IValue>`.
ghstack-source-id: 140695395

Test Plan:
Added automated tests for TupleElements.

Pixel 3 before: https://www.internalfb.com/intern/aibench/details/761596366576284
Pixel 3 after: https://www.internalfb.com/intern/aibench/details/591414145082422
We went from 347 ms to 302 ms.

Reviewed By: dhruvbird

Differential Revision: D30592622

fbshipit-source-id: 93625c54c9dca5f765ef6d5c191944179cb281a8

Author: swolchok

aten/src/ATen/core/ivalue.cpp

@@ -51,19 +51,19 @@ TORCH_API c10::intrusive_ptr<ConstantString> ConstantString::create(
 }
 
 bool operator==(const ivalue::Tuple& lhs, const ivalue::Tuple& rhs) {
-  return lhs.elements_.size() == rhs.elements_.size() &&
+  return lhs.size() == rhs.size() &&
       // see [container equality]
       std::equal(
-             lhs.elements_.cbegin(),
-             lhs.elements_.cend(),
-             rhs.elements_.cbegin(),
+             lhs.elements().cbegin(),
+             lhs.elements().cend(),
+             rhs.elements().cbegin(),
              _fastEqualsForContainer);
 }
 
 TupleTypePtr Tuple::type() const {
   if (!type_) {
     type_ = TupleType::create(
-        fmap(elements_, [&](const IValue& v) { return v.type(); }));
+        fmap(elements(), [&](const IValue& v) { return v.type(); }));
   }
   return type_;
 }