Bug: `to_one_hot()` produces incorrect output for documented `(N, 1)` label arrays

### Description

The `to_one_hot()` utility is documented to accept label arrays of shape `(N,)` or `(N, 1)`.

However, when a valid `(N, 1)` input is provided, the generated one-hot encoding is incorrect due to NumPy broadcasting during advanced indexing.

### File

`deepchem/metrics/metric.py`

### Current Implementation

```python
y_hot[np.arange(N), y.astype(np.int64)] = 1
```

When `y` has shape `(N, 1)`, NumPy broadcasts the indexing arrays instead of performing element-wise indexing, resulting in incorrect assignments across the output matrix.

### Expected Behavior

For:

```python
y = np.array([[0],
[1],
[0]])
```

`to_one_hot(y)` should return:

```python
[[1. 0.]
[0. 1.]
[1. 0.]]
```

### Actual Behavior

The current implementation returns:

```python
[[1. 1.]
[1. 1.]
[1. 1.]]
```

### Impact

This bug affects a documented input shape and can silently generate invalid one-hot encodings.

Potential consequences include:

* Incorrect classification metrics
* Corrupted evaluation results
* Incorrect training targets for classification workflows
* Silent failures that are difficult to detect

### Steps to Reproduce

```python
import numpy as np
from deepchem.metrics import to_one_hot

y = np.array([[0], [1], [0]])

print(to_one_hot(y))
```

Output:

```python
[[1. 1.]
[1. 1.]
[1. 1.]]
```

### Root Cause

The indexing operation:

```python
y_hot[np.arange(N), y.astype(np.int64)] = 1
```

assumes that `y` is one-dimensional.

When `y` has shape `(N, 1)`, NumPy broadcasting changes the indexing semantics and writes to unintended locations in the output array.

### Suggested Fix

Convert `y` to a one-dimensional array before performing advanced indexing:

```python
y = np.asarray(y).reshape(-1)
```

or an equivalent flattening operation before constructing the one-hot representation.

### Additional Notes

The function documentation explicitly states that inputs of shape `(N, 1)` are supported, so the current behavior violates the documented contract.

Start in deepchem/metrics/metric.py and inspect to_one_hot() together with its documented input-shape handling. Reproduce the issue with the provided (N, 1) NumPy label array, then verify that the output matches the expected one-hot matrix and that the existing one-dimensional input behavior remains correct.

Written by the indexing model from the issue text.