Support for custom initial belief (e.g., solving many POMDPs with varying beliefs)

Currently, `_vectorized_initialstate` takes the `pomdp` and `ordered_states` then uses the `initialstate(::POMDP)` to initialize `b0`; however, SARSOP is able to begin from any point in the belief-space, so it makes sense to allow users to provide an initial belief (perhaps in sparse form?) from which to commence SARSOP.

I'd be happy to write/submit a PR for this, but wanted to open an issue to see if it was something y'all would be willing to accept. 🙂

Changes that I think that would be necessary to support this. I could probably tackle this in the next week or two, if an acceptable PR.

```diff
# solver.jl
+ function POMDPTools.solve_info(solver::SARSOPSolver, pomdp::POMDP; b0=initialstate(pomdp))
+ tree = SARSOPTree(solver, pomdp; b0)
- function POMDPTools.solve_info(solver::SARSOPSolver, pomdp::POMDP)
- tree = SARSOPTree(Solver, pomdp)
# the rest of the code ...
return pol, (; ...)
end

+ function POMDPs.solve(solver::SARSOPSolver, pomdp::POMDP; b0=initialstate(pomdp)) =
- function POMDPs.solve(solver::SARSOPSolver, pomdp::POMDP) =
fist(solve_info(solver, pomdp; b0))

# tree.jl
+ function SARSOPTree(solver, pomdp::POMDP; b0=initialstate(pomdp))
- function SARSOPTree(solver, pomdp::POMDP)
+ sparse_pomdp = ModifiedSparseTabular(pomdp, b0)
- sparse_pomdp = ModifiedSparseTabular(pomdp)
# the rest of the codebase ...
return insert_root!(...)
end

# sparse_tabular.jl
+ function ModifiedSparseTabular(pomdp::POMDP, b0)
- function ModifiedSparseTabular(pomdp::POMDP)
S = ordered_states(pomdp)
# the rest of the codebase ...
+ b0 = _vectorized_initialstate(pomdp, S, b0)
- b0 = _vectorized_initialstate(pomdp, S)
return ModifiedSparseTabular(T, R, O, terminal, b0, discount(pomdp))
end

+ function _vectorized_initialstate(pomdp, S, b0)
+ function _vectorized_initialstate(pomdp, S)
b0_vec = Vector{Float64}(undef, length(S))
@inbounds for i ∈ eachindex(S, b0_vec)
b0_vec[i] = pdf(b0, S[i])
end
return sparse(b0_vec)
end
```

So long as `b0` is guaranteed to be compatible with `pdf(b0, S[i])`, these are all the changes necessary. Otherwise, there would need to be some handling to ensure that `b0` is either compatible with vectorizing or allow folks to specify how to achieve the equivalent of `pdf(b0, S[i])`.

---

Perhaps if there were functions in `POMDPTools` that support this kinda interface, that could be interesting but is way out of scope for this issue/PR. (e.g. going from marginal beliefs to relevant `SparseCat` and the like.)