Sync to upstream/release/711

[aatxe]

Hi there, folks! We're back with another weekly Luau release!

Language

• Adds the const keyword for defining constant bindings that are statically forbidden to be reassigned to. This implements luau-lang/rfcs#166.
• Adds a collection of new math constants to Luau's math library per luau-lang/rfcs#169.

Analysis

• Fixes a class of bugs where Luau would not retain reasonable upper or lower bounds on free types, resulting in types snapping to never or unknown despite having bounds.

--!strict
-- `lines` will be inferred to be of `{ string }` now, and prior
-- was 
local lines = {}
table.insert(lines, table.concat({}, ""))
print(table.concat(lines, "\n"))

--!strict
-- `buttons` will be inferred to be of type `{ { a: number } }`
local buttons = {}
table.insert(buttons, { a = 1 })
table.insert(buttons, { a = 2, b = true })
table.insert(buttons, { a = 3 })

• Disables the type error from string.format when called with a dynamically-determined format string (i.e. a non-literal string argument with the type string) in response to user feedback about it being too noisy.
• Resolves an ICE that could occur when type checking curried generic functions. Fixes attempting to modify bounds of a potentially visited generic #2061!
• Fixes false positive type errors from doing equality or inequality against nil when indexing from a table
• In pass a state parameter to useratom #2256, adds a state parameter to the useratom callback for consistency with other callbacks.

Compiler

• Improves the compiler's type inference for vector component access, numerical for loops, function return types and singleton type annotations, fixing [Compiler] things annotated to true / false (?) seem to emit wrong type #2244 MUL & ADD aren't replaced by their constant counterparts in specific scenarios #2235 and MULK bytecode instruction isnt emitted when it should be #2255.

Native Code Generation

• Fixes a bug where some operations on x86_64 would produce integers that would take up more than 32-bits when a 32-bit integer is expected. We resolve these issues by properly truncating to 32-bits in these situations.
• Improves dead store elimination for conditional jumps and fastcalls arguments, improving overall native codegen performance by about 2% on average in benchmarks, with some benchmarks as high as 25%.