[rpc] Fix assertion on vector length during message parsing (pytorch#…

…108414)

Hi!

I've been fuzzing different pytorch modules with with [sydr-fuzz](https://github.com/ispras/oss-sydr-fuzz/tree/master/projects/pytorch), and found a heap buffer overflow error that occurs during Python object deserialization routine. Vector with `IValues` is verified to contain at least 3 elements, which are subsequently removed from vector. The rest of vector is passed further, where it is expected to contain at least one more element. The crash occurs on empty vector.

Docker to reproduce found error: [Dockerfile](https://github.com/ispras/oss-sydr-fuzz/tree/master/projects/pytorch).

### PoC:
[crash-6d634f38a76bfeaa1fffc9472e8ea7b88ee8e776.txt](https://github.com/pytorch/pytorch/files/12499089/crash-6d634f38a76bfeaa1fffc9472e8ea7b88ee8e776.txt)

### ASAN report
```
==339647==ERROR: AddressSanitizer: heap-buffer-overflow on address 0x604000105388 at pc 0x000000c2b3bc bp 0x7fffffffb8d0 sp 0x7fffffffb8c8
READ of size 4 at 0x604000105388 thread T0
    #0 0xc2b3bb in c10::IValue::isString() const /pytorch/aten/src/ATen/core/ivalue.h:685:27
    #1 0xc2b3bb in c10::IValue::toStringRef[abi:cxx11]() const /pytorch/aten/src/ATen/core/ivalue_inl.h:2308:3
    #2 0x101ce65f in torch::distributed::rpc::SerializedPyObj::fromIValues(std::vector<c10::IValue, std::allocator<c10::IValue> >) /pytorch/torch/csrc/distributed/rpc/types.cpp:103:39
    #3 0x1006a7a0 in torch::distributed::rpc::PythonRemoteCall::fromMessage(torch::distributed::rpc::Message const&) /pytorch/torch/csrc/distributed/rpc/python_remote_call.cpp:58:26
    #4 0x101d02e1 in torch::distributed::rpc::deserializeRequest(torch::distributed::rpc::Message const&) /pytorch/torch/csrc/distributed/rpc/utils.cpp:111:14
    #5 0x8db738 in LLVMFuzzerTestOneInput /message_deserialize.cc:192:27
    #6 0x8d84cd in ExecuteFilesOnyByOne /AFLplusplus/utils/aflpp_driver/aflpp_driver.c:255:7
    #7 0x8d82d8 in LLVMFuzzerRunDriver /AFLplusplus/utils/aflpp_driver/aflpp_driver.c
    #8 0x8d7e98 in main /AFLplusplus/utils/aflpp_driver/aflpp_driver.c:300:10
    #9 0x7ffff7a37082 in __libc_start_main (/lib/x86_64-linux-gnu/libc.so.6+0x24082) (BuildId: 1878e6b475720c7c51969e69ab2d276fae6d1dee)
    #10 0x817c4d in _start (/message_deserialize_afl+0x817c4d)

0x604000105388 is located 8 bytes to the left of 48-byte region [0x604000105390,0x6040001053c0)
allocated by thread T0 here:
    #0 0x8d54ca in operator new(unsigned long) /llvm-project-llvmorg-14.0.6/compiler-rt/lib/asan/asan_new_delete.cpp:95:3

SUMMARY: AddressSanitizer: heap-buffer-overflow /pytorch/aten/src/ATen/core/ivalue.h:685:27 in c10::IValue::isString() const
```
Pull Request resolved: pytorch#108414
Approved by: https://github.com/ezyang

torch/csrc/distributed/rpc/python_remote_call.cpp

@@ -46,8 +46,8 @@ std::unique_ptr<PythonRemoteCall> PythonRemoteCall::fromMessage(
 
   // remove the last elements from values and convert it back to an RRef
   TORCH_INTERNAL_ASSERT(
-      values.size() >= 3,
-      "Expect at least 3 elements in the unpickled values, but got ",
+      values.size() > 3,
+      "Expect at least 4 elements in the unpickled values, but got ",
       values.size());
   bool isAsyncExecution = values.back().toBool();
   values.pop_back();