init_inference rejects dtype=torch.int8 on every accelerator

Describe the bug
deepspeed.init_inference(model, dtype=torch.int8) raises ValueError: Data type torch.int8 is not supported by <device> accelerator unconditionally, on every accelerator DeepSpeed ships
(checked cpu, cuda, hpu, mlu, xpu — none list torch.int8 in supported_dtypes()).
This is a regression: two unrelated cleanup PRs (#6528, "add bfloat16 to inference support
dtypes", 2024-09-27; and #5505, "remove unused _validate_args function", 2025-01-07) combined to
silently tighten InferenceEngine.__init__'s dtype gate (deepspeed/inference/engine.py:74) from
its original fp16-only check to a blanket supported_dtypes() check that was never meant to
describe DeepSpeed's own int8 weight-quantization dtype. docs/_tutorials/inference-tutorial.md
("Datatypes and Quantized Models", rewritten as recently as #8535) still documents int8 as
supported.

Found while investigating #1454 — the compute_attention()/input_mask crash reported there is
already fixed upstream, but the underlying "run int8 quantized inference" workflow that issue was
trying to use is broken by this separate, still-live gap.

To Reproduce

import torch, deepspeed
deepspeed.init_inference(torch.nn.Linear(4, 4), dtype=torch.int8)

No GPU needed — the gate itself is device-agnostic and raises identically on CPU.

Expected behavior
init_inference should accept dtype=torch.int8 as it did before #6528, since DeepSpeed's own
int8 weight-quantization path (DeepSpeedSelfAttention/DeepSpeedMLP in
deepspeed/ops/transformer/inference/ds_attention.py) still special-cases torch.int8
construction today — that code is presently unreachable dead code because the engine can no
longer be constructed with this dtype at all.

System info:

• Reproduces identically on every accelerator backend (cpu, cuda, hpu, mlu, xpu) since
  the dtype check is evaluated before any accelerator-specific code runs.
• Python 3.12, torch 2.14 (CPU repro); not version-specific — same gate on current master.

Additional context
A one-line fix (exempt torch.int8 from the supported_dtypes() gate, matching the code's
pre-#6528 scope) plus a regression test is ready; opening a PR alongside this issue.

Start at deepspeed/inference/engine.py:74 and compare the dtype gate with the int8 handling in deepspeed/ops/transformer/inference/ds_attention.py. Reproduce the failure with the provided CPU example, then add a regression test for init_inference with torch.int8. Done means int8 initialization is accepted and the regression test passes.