…rations (pytorch#155544)

Improves the GEMM overview logging in PyTorch Inductor to properly display batch size information for batched matrix operations like `torch.bmm` and `torch.baddbmm`.

**Fixes pytorch#155307**

## Problem

The current GEMM logging for `torch.bmm` shows:
```python
# Repro
import os
os.environ["TORCH_LOGS"] = "inductor"
import torch

M, N, K = 1024, 1024, 1024
dtype = torch.bfloat16
A = torch.randn(10, M, K, device="cuda", dtype=dtype)
B = torch.randn(10, K, N, device="cuda", dtype=dtype)

compiled_model = torch.compile(torch.bmm, fullgraph=True)
_ = compiled_model(A, B)
```

**Before:**
```
Name                 | M                    | N                    | K                    | Count
----------------------------------------------------------------------------------------------------
aten.bmm             | 1024                 | 1024                 | 1024                 | 1
----------------------------------------------------------------------------------------------------
```

The batch size (10) is missing from the logs, making it unclear what the actual operation dimensions were.

## Solution

**After:**
```
Name                           | B                    | M                    | N                    | K                    | Count
----------------------------------------------------------------------------------------------------------------------------------
aten.bmm                      | 10                   | 1024                 | 1024                 | 1024                 | 1
aten.mm                       | -                    | 1024                 | 1024                 | 1024                 | 2
----------------------------------------------------------------------------------------------------------------------------------
```

## Changes Made

### 1. Enhanced Parsing Logic in compile_fx.py
- Detects batched operations by checking if operation name ends with `'bmm'` or `'baddbmm'`
- For batched operations: takes last 4 parts as `batch, m, n, k`
- For non-batched operations: takes last 3 parts as `m, n, k`
- **Dedicated "B" column**: Added separate column for batch size instead of embedding in operation name
- Shows batch size for batched operations, shows "-" for non-batched operations

### 2. Updated All MM Operations for Consistency
- **bmm.py**:
  - Extract batch size from `mat1.get_size()[0]` for both `tuned_bmm` and `tuned_baddbmm`
  - Use positional counter keys: `aten.bmm_{batch_size}_{m}_{n}_{k}`
  - Enhanced log messages to include batch size information

- **mm.py**: Updated counter keys for consistency:
  - `aten.mm_{m}_{n}_{k}` (no batch dimension)
  - `aten.addmm_{m}_{n}_{k}` (no batch dimension)
  - `aten._int_mm_{m}_{n}_{k}` (no batch dimension)
  - `aten._scaled_mm.default_{m}_{n}_{k}` (no batch dimension)

Pull Request resolved: pytorch#155544
Approved by: https://github.com/jansel, https://github.com/BoyuanFeng