fix(caching): use dtype.itemsize instead of finfo().bits for byte size

[Gpgabriel25]

Problem

jnp.finfo(dtype).bits // 8 crashes with ZeroDivisionError for dtypes
that lack an finfo entry — float4_e2m1fn, integer types, and other
exotic quantization formats. This silently blocks use of low-precision
KV caches in UnifiedAttentionCacheConfig, RaggedPagesCacheConfig,
MlaRaggedPagesCacheConfig, and get_dtype_size() in the eSurge engine.

Fix

Replace all five occurrences with jnp.dtype(dtype).itemsize, which is
the correct, dtype-agnostic way to query byte width in JAX.

Evidence

Probe run against main vs this branch, same venv, jnp.float4_e2m1fn:

branch	unified num_pages	ragged num_pages	ratio fp4/bf16
main	ZeroDivisionError	ZeroDivisionError	—
this PR	bf16: 262 144 → fp4: 524 288	bf16: 262 144 → fp4: 524 288	2.0×

fp4 now correctly gets double the page capacity relative to bf16, as
expected from halving the per-element byte cost.

Files

• easydel/caching/_specs.py
• easydel/caching/mla_ragged_page/cache.py
• `easydel/caching/ragged_paBody:

## Problem

`jnp.finfo(dtype).bits // 8` crashes with `ZeroDivisionError` for dtypes
that lack an `finfo` entry — `float4_e2m1fn`, integer types, and other
exotic quantization formats. This silently blocks use of low-precision
KV caches in `UnifiedAttentionCacheConfig`, `RaggedPagesCacheConfig`,
`MlaRaggedPagesCacheConfig`, and `get_dtype_size()` in the eSurge engine.

## Fix

Replace all five occurrences with `jnp.dtype(dtype).itemsize`, which is
the correct, dtype-agnostic way to query byte width in JAX.

## Evidence

Probe run against `main` vs this branch, same venv, `jnp.float4_e2m1fn`:

| branch | unified num_pages | ragged num_pages | ratio fp4/bf16 |
|--------|------------------|-----------------|----------------|
| main   | `ZeroDivisionError` | `ZeroDivisionError` | — |
| **this PR** | bf16: 262 144 → fp4: 524 288 | bf16: 262 144 → fp4: 524 288 | **2.0×** |

fp4 now correctly gets double the page capacity relative to bf16, as
expected from halving the per-element byte cost.

## Files

- `easydel/caching/_specs.py`
- `easydel/caching/mla_ragged_page/cache.py`
- `easydel/caching/ragged_page/cache.py`
- `easydel/caching/unified_attention/cache.py`
- `easydel/inference/esurge/utils.py`

## Tests

- `tests/inference/test_esurge_kvdtype_override.py` (new, 5 tests)
- `tests/layers/caching/test_kv_budget_parallelism.py` (new, 2 tests)
```ge/cache.py`
- `easydel/caching/unified_attention/cache.py`
- `easydel/inference/esurge/utils.py`

## Tests

- `tests/inference/test_esurge_kvdtype_override.py` (new, 5 tests)
- `tests/layers/caching/test_kv_budget_parallelism.py` (new, 2 tests)