add device attribute in ETensor

[Gasoonjia]

[Gasoonjia]

We've aligned on the design;
The work will be broken into two phases:

• Phase 1: End-to-end device type support without user-specified device placement
  We’ll make the E2E pipeline work, including new device type, memory planing for input and output tensors, emitter, tensor_parser, etc.
  We'll rely on the partitioner to determine device type of tensors: tensors in the partitioned graphes and their input and output tensors will be on target device, and rest will be still on cpu.
  The CUDA backend will serve as the reference implementation, exercising the new DeviceAllocator interface, schema changes, and runtime APIs.

• Phase 2: Custom device placement support
  We'll provide an interface that lets users explicitly specify whether a subgraph should be lowered to a particular device.
  Before starting this phase, we'll share a few proposed options to gather feedback from the group.

[Gasoonjia]

#17533

[Gasoonjia]

@JacobSzwejbka made out a good point that current singleton planed-buffer owned allocator may introduce memory conflict between different method. Tweaking the design a little bit to make sure the planed-device-buffer is owned by runtime/method

[Gasoonjia]

#18078 and beyond

[Gasoonjia]

Right now we have support device attribute annotation support, as well as non-cpu memory planing support as well. Working on the some bugs here and will working on the graph functionality, then will try to do some benchmark.

[Gasoonjia]

we've inserted the h2d and d2h copy in the graph to make graph functional #18730

[Gasoonjia]

#18788
we have e2e work on qwen3.5 model