Performance bottlenecks in mesa.discrete_space

**Summary**

While profiling mesa.discrete_space under a high-movement workload, I observed significant time spent in cell movement and occupancy checks. This issue summarizes the hotspots and proposes a few improvements.

**Reproduction**
````python
import random
import cProfile
from mesa import Model
from mesa.discrete_space.grid import OrthogonalVonNeumannGrid
from mesa.discrete_space.cell_agent import CellAgent

class DummyModel(Model):
pass

def run_benchmark():
model = DummyModel()
grid = OrthogonalVonNeumannGrid((100, 100), capacity=1)

agents = [CellAgent(model) for _ in range(5000)]

# initial placement
for agent in agents:
agent.cell = grid.select_random_empty_cell()

# random movement loop
for _ in range(500):
for agent in agents:
neighbors = agent.cell.neighborhood
target = random.choice(list(neighbors))
if target.is_empty:
agent.cell = target

if __name__ == "__main__":
cProfile.run("run_benchmark()", sort="cumtime")
````
**Observations**

- `Cell.is_empty` is called very frequently and currently computes `len(self._agents) == 0` every time.
- `CellAgent.cell` triggers add or remove operations that account for a large portion of total runtime.
- Neighborhood access wraps results in `CellCollection` and constructs intermediate mappings, contributing additional allocation overhead.

Total runtime is around 63 seconds
````
CellAgent.cell ~14.7s
Cell.is_empty ~11.1s
Cell.neighborhood ~6.8s
Cell.remove_agent ~6.1s
Cell.add_agent ~4.8s
````
**Proposed Improvements**

- Make `_empty` the single source of truth for emptiness and avoid recomputing `len(self._agents)` on every `is_empty` call.
- Reduce overhead in the movement pathnto avoid unnecessary repeated checks and work.
- Simplify how neighborhoods are constructed and returned to reduce intermediate allocations during frequent movement.

I plan to address these in separate PRs to keep changes focused and measurable.

Start by running the provided cProfile benchmark, then inspect Cell.is_empty, CellAgent.cell, Cell.add_agent, Cell.remove_agent, and neighborhood access in mesa.discrete_space. Treat the proposed changes as separate, measurable efforts; done means reducing the identified hotspots without changing movement or occupancy behavior.

Written by the indexing model from the issue text.